UNIVERSITY OF OKLAHOMA

GRADUATE COLLEGE

MISINFORMATION DETECTION METHODS USING LARGE LANGUAGE MODELS

AND EVALUATION OF APPLICATION PROGRAMMING INTERFACES

A THESIS

SUBMITTED TO THE GRADUATE FACULTY

in partial fulfillment of the requirements for the

Degree of

MASTER OF SCIENCE

By

CATHERINE G. G. DONNER
Norman, Oklahoma
2024
MISINFORMATION DETECTION METHODS USING LARGE LANGUAGE MODELS

AND EVALUATION OF APPLICATION PROGRAMMING INTERFACES

A THESIS APPROVED FOR THE

GALLOGLY COLLEGE OF ENGINEERING

BY THE COMMITTEE CONSISTING OF

Dr. David Ebert, Chair

Dr. Dean Hougen

Dr. Katerina Tsetsura

Dr. Jeong-Nam Kim

Dr. Naveen Kumar

© Copyright by CATHERINE G. G. DONNER 2024
All Rights Reserved.
Contents

Abstract ix

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Significance to Data Science Field . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Significance to Society and Stakeholders . . . . . . . . . . . . . . . . . . . . 7

2 Literature Review 9
2.1 Research Problem Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Applications to Stakeholders . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Misinformation from Communications Perspective . . . . . . . . . . . . . . . 12
2.3.1 Concepts Behind Natural Language Processing Dimensions . . . . . . 14
2.4 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Large Language Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5.1 Theory Behind Supervised Learning and Transformer Model . . . . . 17
2.5.2 Bidirectional Encoder Representations from Transformers . . . . . . . 19
2.5.3 Robustly Optimized BERT Approach . . . . . . . . . . . . . . . . . . 20
2.5.4 Distilled BERT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5.5 Applications of Large Language Models to Natural Language
Processing Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

iv
 2.5.6 Applications of Large Language Models to Application
Programming Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6 Unsupervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6.1 Theory Behind Unsupervised Learning . . . . . . . . . . . . . . . . . 23
2.6.2 Association Rule Mining . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6.3 Anomaly Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.6.4 Cosine Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.6.5 t-Distributed Stochastic Neighbor Embedding . . . . . . . . . . . . . 25
2.6.6 Latent Dirichlet Allocation Modeling . . . . . . . . . . . . . . . . . . 25
2.6.7 Applications of Unsupervised Learning to
Misinformation Detection . . . . . . . . . . . . . . . . . . . . . . . . 26
2.7 Generative Artificial Intelligence Models . . . . . . . . . . . . . . . . . . . . 27
2.7.1 Theory Behind Artificial Intelligence . . . . . . . . . . . . . . . . . . 27
2.7.2 OpenAI ChatGPT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.7.3 Google PaLM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.7.4 Applications of Using AI Models for Misinformation
Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.8 Research Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Supervised Learning Using Large Language

Models and Natural Language Processing Dimensions 32
3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 Natural Language Processing Dimensions . . . . . . . . . . . . . . . . . . . . 33
3.3 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.4 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5 Exploratory Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5.1 Article Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

v
 3.5.2 Article Title . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.5.3 Article Origin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5.4 Natural Language Processing Dimensions . . . . . . . . . . . . . . . . 48
3.6 Data Labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.7 Model Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.7.1 BERT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.7.2 RoBERTa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.7.3 DistilBERT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.8 Results and Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.9 Model Integration in Proposed Data Tool . . . . . . . . . . . . . . . . . . . . 61
3.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 Unsupervised Learning Insights and Data Visualizations 65

4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3 Method 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.1 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.2 Anomaly Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.3 Latent Dirichlet Allocation Topic Modeling . . . . . . . . . . . . . . . 68
4.3.4 Results and Interpretation . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 Method 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4.1 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4.2 Cosine Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4.3 t-Distributed Stochastic Neighbor Embedding . . . . . . . . . . . . . 74
4.4.4 Results and Interpretation . . . . . . . . . . . . . . . . . . . . . . . . 75
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.6 Note on Association Rule Mining and Related Work . . . . . . . . . . . . . . 79

vi
5 Evaluation of Artificial Intelligence Model
Usage for Misinformation Detection 80
5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.2 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.3 Application Programming Interface Setup . . . . . . . . . . . . . . . . . . . 82
5.3.1 ChatGPT-3.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3.2 ChatGPT-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3.3 PaLM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.4 Comparative Analysis and Interpretation . . . . . . . . . . . . . . . . . . . . 85
5.4.1 Reasons for Scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.4.2 Accuracy in Scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

6 Discussion 98
6.1 Implication of Findings to Research Problems . . . . . . . . . . . . . . . . . 98
6.1.1 Data Tool Using Large Language Model . . . . . . . . . . . . . . . . 98
6.1.2 Unsupervised Learning Approaches . . . . . . . . . . . . . . . . . . . 100
6.1.3 Generative Artificial Intelligence Model Evaluation . . . . . . . . . . 102
6.1.4 The Most Efficient Model . . . . . . . . . . . . . . . . . . . . . . . . 103
6.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.3 Potential for Improvement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.4 Societal Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.5 Summary of Limitations and Improvements . . . . . . . . . . . . . . . . . . 110
6.6 Note on Subjectivity of Misinformation . . . . . . . . . . . . . . . . . . . . . 111

7 Conclusion 113
7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.2 Contribution to Data Science Field . . . . . . . . . . . . . . . . . . . . . . . 114

vii
 7.3 Contribution to Society and Stakeholders . . . . . . . . . . . . . . . . . . . . 115
7.4 Future Work and Recommendations . . . . . . . . . . . . . . . . . . . . . . . 116
7.5 Final Thoughts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

8 Acknowledgements 120

9 References 122

10 Appendices 133
10.1 Appendix A: Supplementary Code Snippets . . . . . . . . . . . . . . . . . . 133
10.1.1 NLP Dimension Vocabularies . . . . . . . . . . . . . . . . . . . . . . 133
10.1.2 Data Tool Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
10.2 Appendix B: Additional Visualizations and Images . . . . . . . . . . . . . . 141
10.2.1 LLM Training Versus Validation . . . . . . . . . . . . . . . . . . . . . 141
10.2.2 Association Rule Mining on Politifact Kaggle Dataset . . . . . . . . . 144
10.3 Appendix C: Glossary of Commonly Used Terms and Acronyms . . . . . . . 149

viii
Abstract

Misinformation has emerged as a pressing public policy concern, prompting transdisciplinary

research in the data science field. News journalism provides a foundation for free speech in
modern society, yet misinformation in mainstream and independent media through opin-
ionated or biased news can pose dangerous consequences ranging from misunderstanding
basic facts to emboldened extremism. Currently, the preeminent tool for misinformation
detection is the large language model (LLM) as it is renowned for its ability to capture
the context and meaning of textual data. In addition, generative artificial intelligence (AI)
models, namely OpenAI’s ChatGPT and Google’s Pathways Language Model (PaLM), are
accessible in an application programming interface (API) form, which can also provide op-
portunities for automated misinformation detection. Despite advancements in developing
effective data science tools for identifying misinformation, there are not many available op-
tions, and it is crucial to assess pre-existing tools to determine the most recommended model
to pursue for future field research and open-source use in automated misinformation detec-
tion. This thesis attempts to evaluate fine-tuned supervised LLMs, AI model frameworks,
and unsupervised learning methods to propose an explainable, automated misinformation
detection tool that incorporates multiple natural language processing (NLP) dimensions and
holistically evaluates trustworthiness in news articles. The study revealed that the Hugging
Face LLM RoBERTa with added NLP dimensions as features was the most effective model.
Furthermore, it was found that unsupervised learning methods provided valuable insights
that eliminated some ambiguity between trustworthy and fake news articles, and AI models

ix
tended to inflate the trustworthiness values of news articles.

Keywords: misinformation, large language models (LLMs), unsupervised learning, appli-

cation programming interfaces (APIs)

x
Chapter 1

Introduction

1.1 Background

To begin, the term misinformation is defined as an incorrect or misleading statement that can
obscure the truth [29]. A similar infodemic, disinformation, is defined as false information
deliberately created or spread in order to cause harm, usually with political, psychological,
or social motivations [21]. Misinformation (along with disinformation) is a rapidly growing
problem in the modern world today, and while ideally widespread media literacy and regu-
lations on social media content are the most effective methods for helping stop the spread of
misinformation [15], there is more demand for taking on massive amounts of misinformation
using big data analytical tools through fast and accurate classification of misinformation.
Data science, as a STEM field with multiple interdisciplinary applications, can have the
technological ability to attain this difficult objective. However, to do this it is important
to understand which contributions are the most effective for scientific knowledge and soci-
etal purposes, whether these be through the creation of new tools, or the discovery of new
data insights, on certain aspects of the misinformation domain. This thesis attempts to
quantitatively examine aspects of misinformation that are visible (trustworthiness based on
biased news) and invisible (revealing hidden patterns among trustworthy and untrustworthy
articles).

This thesis was inspired as a result of issues discovered through the course of an assistantship

1
project. The objective of this project was to devise a framework for creating trustworthiness
scores based on certain dimensions of article texts. It was a requirement that databases of
information were to be populated with the scores at an industry-level accelerated pace to
meet deadlines, and APIs were used to populate prior fields of information. An API is a
type of interface that connects a computer to a server to return information about an input
entity. For example, the Google Translate API provides the translation for a given text and
target language [55]. APIs are very efficient at providing multi-faceted information about an
entity with a quick response time, making them suitable for programming tasks in industry.

Since it was preferable to use an API for populating trustworthiness scores for the database,
an option that was considered was the Romanian-based Zetta Cloud TrustServista API,
which provided misinformation scores as well as explanations for what dimensions were
considered for the scores (i.e., named entities, clickbait, sentiment). However, due to its lack
of transparency on cost, this API was not considered and a machine learning approach was
eventually used. Besides the TrustServista API, very few APIs were available that could
perform the intended task of wholesomely determining a trustworthiness score for any given
text, regardless of length. Thus, the lack of availability of efficient automated misinformation
detection tools made goals in this project more difficult to accomplish, and the idea behind
this thesis was conceived.

1.2 Research Questions

Given the context of the problem presented in the previous section, pressing research ques-
tions arise. Additionally, research questions involving the general efficiency of pre-existing
tools need to be addressed.

Can a data science tool that rapidly and accurately assesses the misinformation
likelihood of news articles, while ensuring transparency through incorporating
key NLP dimensions, be developed to assist stakeholders who are potentially

2
impacted by, or may have a role in, combating misinformation?

Given that there are not many tools available for evaluating the overall trustworthiness of a
news source [52], more options should be added to the current competition to have a variety
of tools for misinformation detection tasks done on different types of media, especially when
it comes to analyzing news media. Creating a data tool that can provide a trustworthiness
score for a given news article text, preferably on a scale from 0 to 100 rather than on a binary
scale, as well as explain why it gave the text that score could potentially be an effective and
explainable misinformation detection tool. To explain the prediction of the trustworthiness
score, quantified NLP dimensions will be incorporated and should provide a suitable frame-
work for integrating dimensions of misinformation related to the field of communications,
contributing to transdisciplinary work. For the base model of this tool, an LLM will be used
as it is currently one of the most effective models for executing NLP tasks; multiple base
LLMs will be trained and compared for performance. Misinformation detection is usually a
classification-based task, however for the LLMs this problem will be a regression problem as
the scale from 0 to 100 provides an opportunity to predict the true likelihood of a news article
being misinformation instead of denoting by a label (i.e., 0 or 1) that does not have much
middle ground in terms of the real likelihood that a news article is trustworthy. In addition,
usage of this tool by the target stakeholders could help reduce the spread of misinformation
and its harmful effects on populations, making the tool a potentially valuable contribution
to information warfare.

Which dimensions of misinformation will be considered when creating the data

science tool for scoring the trustworthiness of a news article?

Besides data science being a field with major contributions to evaluating misinformation
potential, misinformation is highly researched in the field of communications as it is largely
based on the communicative meaning between words and phrases present as well as the re-
lationship between the author who disseminates the information and the users who consume

3
that information [10], [11], [12], [43]; these meanings can be capable of being converted into
NLP interpretations. 13 dimensions have been proposed for this thesis based on literature
review [3], [5], [16], [26], [40], [43], [50], [58], [82] and distribution variability: sentiment,
persuasion, exaggeration, context, inclusion of multiple perspectives, use of named entities,
use of statistics, referencing of previous articles, term frequency, distraction, verification of
claims, logical coherence, and clickbait title. Specific features of these dimensions will be
covered in Chapter 2. These dimensions are most indicative of biased or opinionated news
sources, which can have a greater likelihood of spreading misinformation and being less trust-
worthy. These dimensions will also help provide explainability for the proposed data science
tool and will be converted into quantitative inputs for the tool using specified vocabulary
sets. When it comes to the cumulative quality of these proposed dimensions, these are di-
mensions that should be universally considered when determining if a given news article is
likely to be misinformation as these should not overlap too much in features but provide
sufficiently broad accountability in determining trustworthiness.

Can unsupervised learning methods performed on trustworthy and untrustwor-

thy news articles provide any valuable insights for future research on misinfor-
mation detection?

Misinformation when presented in an objective tone can pose an invisible but dangerous
risk when not held accountable, and the use of unsupervised learning methods has the po-
tential to unveil patterns among this type of misinformation. Although some work may be
completed later using a verified true and false statement dataset that can provide contribu-
tions to future research on automated fact-checking, this thesis attempts to analyze patterns
between the articles that can also contribute to visualizations using misinformation data.
These insights can additionally be used in training supervised learning models for improving
misinformation detection. Methods including association rule mining, anomaly detection,
cosine similarity, t-Distributed Stochastic Neighbor Embedding (t-SNE), and Latent Dirich-

4
let Allocation (LDA) modeling will be explored in this thesis [9], [32], [39], [53], [73].

Do generative AI models have the potential to detect misinformation accurately?

Relating to the lack of availability of tools like APIs that perform trustworthiness scoring,
the options that exist mainly adapt the frameworks of generative AI models, most notably
OpenAI’s ChatGPT and Google’s PaLM, which is the model framework used in Google Bard.
AI models such as these are increasingly becoming popular due to their user-friendly nature
and ability to inform the user about almost any aspect of existence. However, should they be
relied on for misinformation detection tasks, which can have major risks and consequences
for society if they are not accurate? These AI models in their API forms will be explored and
evaluated in their abilities to rate the trustworthiness of a news article text. These models
have the potential to produce similar scores (from 0 to 100) when prompted by the user that
can then be compared with the labeled and model scores that were calculated using the NLP
dimensions. Responses given by the models appear to be detailed and well-written, including
consideration of multiple dimensions to explain why the model gave the news article text
the score that it gave. However, despite the seemingly efficient and accurate appearance of
the AI models’ ability to detect misinformation, it is important to note that using an AI
model as an open-source tool to detect misinformation should not be taken for granted. It
likewise should be noted that AI models even have the potential to spread misinformation,
so there could, ironically, be misinformation in their ability to detect misinformation. There-
fore, these models should be evaluated, and this thesis can serve as a preliminary study of
how trustworthy these models are for misinformation detection purposes. The AI models’
scores and explanations will be evaluated with the proposed model’s results using various
comparative methods.

1.3 Objectives

There are 3 primary objectives, or components, of this thesis:

5
 1. Component 1: Creating and proposing a data science tool that utilizes an LLM as the
model foundation and NLP dimensions as supporting explainability and prediction fea-
tures. This tool will use news article texts as input and will output cumulative trustwor-
thiness scores on a scale from 0 to 100 (0 meaning not trustworthy and high misinfor-
mation potential and 100 meaning very trustworthy and low misinformation potential).
Multiple base LLMs from the Hugging Face website (https://huggingface.co/) in-
cluding BERT, RoBERTa, and DistilBERT [18], [42], [59], [80] will be trained and
compared, and the best performing model that results from this comparison will be
used in the tool. A compiled training dataset of news article texts, titles, and origins
will be used for this component of the thesis.

2. Component 2: Analyzing data patterns between likely trustworthy and likely un-
trustworthy news articles to discover invisible data patterns using unsupervised learn-
ing methods. A separate dataset can be used using true and false statements to elim-
inate the ambiguity of objective misinformation, however for consistency a subset of
the news article dataset will be used for analysis.

3. Component 3: Evaluating generative AI models’ ability to detect misinformation

and score trustworthiness of news article texts and comparing the AI-generated scores
and explanations with the true scores. The AI models that will be explored will be
ChatGPT-3.5, ChatGPT-4, and PaLM in their API forms. The trustworthiness scores
generated from the APIs will be collected and stored in the dataset used for Component
1 since these APIs will be using the same article texts in the training dataset as inputs.

1.4 Significance to Data Science Field

The significance of this thesis to the data science field and general scientific knowledge will
help close some important gaps in the research area of automated misinformation detection.

6
First, Component 1 will help contribute to the creation of a new data science tool to the
current competition of available tools used for predicting trustworthiness (or misinformation
potential) scores for news article texts. This tool can likewise be open to peer review by
experts and professionals in the field who have an interest in creating and evaluating auto-
mated misinformation detection tools.

Next, Component 2 will help generate valuable insights from the unsupervised learning meth-
ods that can provide significant contributions to the field. The results can have the potential
to contribute to future field research on automated fact-checking, improving misinformation
detection models, and predicting trustworthiness scoring for news articles. In addition, data
visualizations will be made to visualize the patterns detected among the trustworthy and
untrustworthy news articles, which can serve as contributions to scientific knowledge about
misinformation data visualizations.

Finally, the results from Component 3 should serve as a preliminary study of how reliable
or unreliable AI tools can be for misinformation detection tasks, and it is likely that more
preliminary studies will come out researching this issue in the coming years.

1.5 Significance to Society and Stakeholders

Besides this thesis’ potential impact on scientific knowledge, an impact on society and target
stakeholders should also be recognized. Listed below are the target societal stakeholders and
how this thesis can benefit them:

• Military agencies (i.e., U.S. Air Force): As the COVID-19 pandemic had a significant
impact on supply chains across the U.S., supply chain operations can also impact the
military, which has a crucial role in ensuring military readiness and national security
interests. Military supply chains can also be impacted by misinformation; for example,
if the military were to do business with a company to request products from and there

7
 is misinformation that is spread about the company that inadvertently harms the
company’s reputation as a potential supplier, the military could refuse to do business
with the company based on that false information. Therefore, the military can benefit
from this research as it can have the potential to help keep supply chain operations
running smoothly by holding misinformation accountable that can sabotage business
decisions.

• Government agencies (i.e., DoD, DHS, CIA, NSA): Various government agencies can
benefit from this thesis as they too have a role in protecting national security interests
and the safety of U.S. citizens. U.S. citizens and government agencies can be subject to
damaging misinformation that can affect livelihoods and reputations, so it is important
for these agencies to utilize efficient data science tools for automated misinformation
detection and to invest in information warfare. A data science tool like the one proposed
in Component 1 of this thesis could potentially later be used by the U.S. government
to combat and tackle misinformation on a big-data scale.

• Programmers in industry: Industry programmers whose work may involve misinforma-

tion detection of texts could be impacted by this research. Industry programmers may
work with tools like APIs to populate databases at an accelerated pace to complete
deliverable deadlines, and having an API-like tool like the one proposed in Component
1 can be an example of a tool that industry programmers can use for accurate and
quick misinformation detection for large databases of texts.

• The public: The tool proposed in Component 1 can also be used for open-source use
by U.S. citizens who want to be informed on whether and why news articles that are
found on mainstream news sources, independent news sources, or social media are
trustworthy or not trustworthy.

8
Chapter 2

Literature Review

2.1 Research Problem Context

As stated in Chapter 1 of this thesis, widespread social media regulations and media literacy
are ideally the most effective methods for mitigating the infodemic of misinformation [15].
A study on the impact of fake news on individuals’ partisanship focusing on third-person
perspectives revealed the presence of unbiased support for interventions aimed at curbing
the spread of misleading information [31].

There are factors that need to be considered when it comes to the news content and the
person absorbing that news content. Fake news is less than 1/10th of 1% of all media
consumption [79], but despite this figure of fake news consumption appearing quite small,
fake news consumption can still be amplified in different contexts. About 1 in 4 Americans
have visited a fake news website, and around 6 out of 10 visits to a certain fake news website
were from the 10% most conservative Americans [23]. The social media network Facebook
has been known to aid in the spread of misinformation [15], [23]. Although the concept of
fact-checking appears to be an effective way to discern fake news from real news, on social
media networks like Facebook this approach does not reach a wide audience of fake news
consumers [23]. Fact-checking has also been shown to not be as effective in mitigating the
risk of misinformation spread and information polarization [81]. Additionally, personality
factors of people consuming fake news include bias, agreeableness, extraversion, negative

9
emotionality, open-mindedness, and hours spent consuming news; these factors can lead to
media consumers’ susceptibility to fake news [6]. Fake news, as it is likewise appealing in
attention-grabbing tactics, has also been shown to increase dopamine levels in people’s brains
during consumption [22], thereby creating an emotional reward for the reader consuming the
news content.

It is also important to note which exact definition and form of misinformation needs to be
considered in the context of the research problems of the thesis. First, there are specific
distinctions between the following infodemic definitions [21]:

• Propaganda: Information shared by the government usually with a political connota-

tion.

• Disinformation: False information that is intentionally disseminated to cause harm.

• Misinformation: False information that is not intentional in harm.

• Malinformation: False information that is deliberately distributed to cause harm to a

person’s or organization’s reputation.

There are yet other types of fake news, including satire, hoax, clickbait, and conspiracy [46],
as well as rumors and spam [29]. Opinionated news will be the target category for Com-
ponent 1 analysis, as this is a form of intentional disinformation and is likely to have more
easily recognizable NLP dimensions to embed including high sentiment, high exaggeration,
and lack of context that can help predict the trustworthiness scores. For analysis involving a
true and false statement dataset, this dataset will contain objective and unintentional mis-
information that will be harder to distinguish from factual statements.

The impact of misinformation has shown evidence of being far-reaching and having major
negative impacts on society and democracy. Social media usage and fake news absorption
have led to increased polarization and misunderstanding in social interactions [4], [70]. Mis-
information likewise poses a threat to the institution of journalism as principles of veracity

10
and objectivity can be violated with the spread of misinformation through the news [81].
One of the most dangerous implications of disinformation, or intentional misinformation, is
the use of false pretenses to target people. This can occur in the case of targeting ethnic
groups, individuals, or countries through extremist ideologies [20]. Fake news can also kill
through pandemics; when inaccurate or downplayed information is being spread in regards
to an encroaching pandemic, this can lead to people not taking preventative measures from
catching the disease [10]. A final example is the ever-growing problem of well-organized,
weaponized disinformation surrounding events of publicity including presidential elections,
vaccination, and even infrastructure networks [30]. These harmful campaigns can have the
potential to affect public services that citizens are dependent on, and disinformation that
can lead to a shutdown of those services can greatly affect supply chain management [2],
[30]. Ensuring that supply chain operations, especially for the benefit of the U.S. military,
are unaffected by this type of disinformation due to highly efficient misinformation detection
has become an important application of this research based on the impacts laid out in this
literature review.

Therefore, especially in regards to the double threat of people’s susceptibility to fake news
and fake news’ potential to harm people, it is important to hold misinformation accountable
and reduce its spread on a big-data scale, through quick and effective automated means [49].
The longer that the issue of misinformation metastasizes, the more the spread of misinfor-
mation becomes out of control, and the less able people can be at distinguishing between
fake and real news.

2.2 Applications to Stakeholders

The intended stakeholders of U.S. government and military agencies, industry programmers,
and public users in which this thesis work could impact involve many applications. First,
when it comes to the overall role of the U.S. military in providing defense for citizens, it

11
is important that military agencies, especially the Air Force, can be able to meet supply
and demand needs to be able to make clear business decisions when building essential mil-
itary equipment. Thus, misinformation can impact supply chain risk management when
false information about a company or country can jeopardize business decisions [2]. Next,
when it comes to government agencies, this thesis work could be applied to certain agencies
(i.e., NSA, CIA) that focus on defending national security interests against misinformation
threats, both foreign and domestic [21], [81].

In the context of programming in industry, not much research has been conducted on misin-
formation tools helping data scientists in industry, however the accessibility of open-source
models and tools such as APIs have been very helpful for programmers in industry to quickly
request information about databases of entities on a large-data scale [83]. This thesis work
intends to benefit this target stakeholder by providing another easy-to-use tool for misinfor-
mation detection purposes. This application is also based on the original background from
Chapter 1 that outlined the inspiration for this thesis. Finally, it is important for public
users to have knowledge of and access to tools that can verify whether the news that they
see on social media, mainstream media, or independent media is trustworthy or not; current
open-source fact-checking applications include Politifact, Gossip Cop, B.S. Detector, and
Fake News Detector AI [63], [81].

2.3 Misinformation from Communications Perspective

Data science is a field that has the potential to tackle the problem of misinformation from
a STEM point of view, however the domain of misinformation can overlap with other aca-
demic fields including communications and journalism. Thus, it is important to consider
the communication dimensions of misinformation and incorporate them into data modeling
through transdisciplinary research.

Numerous dimensions of misinformation exist, both text-based and external. For example,

12
the use of certain pronouns, politeness, emotion, complex vocabulary, and words indicating
uncertainty can mean the difference between fake news and real news [10]. Characterizations
of a higher likelihood of misinformation can include a distant speaker-audience relationship,
high emotion, a goal to illicit audience action, detailed information, and a high number of
participants and sources cited [11]. Factors including diffusion scope (how broad the audi-
ence is), speed (number of users reacting within a certain timeframe), and shape (broadcast
or human-human transmission) have been used to determine misinformation through ana-
lyzing clusters of users reacting to certain tweets on the social media network X, formerly
known as Twitter [12]. Sentiment is possibly the most important textual dimension of mis-
information, and the use of emotion in misinformation can lead viewers to be more engaged
and susceptible to misleading messages [43].

So how can communication dimensions be incorporated into a data science model? This
is through the textual analysis technique of NLP, which is where a computer is trained to
interpret human language; this is usually done through converting words into numerical rep-
resentations, also known as word embedding [76]. NLP tasks such as stance detection, rumor
detection, and sentiment analysis can be performed by converting certain word vocabularies
into numeric inputs [66].

Regarding misinformation detection APIs that are restricted in availability, these can have
the ability to not only provide a trustworthiness score, but also take into account which
communication dimensions were taken into consideration when reading an article text. The
Romanian-based TrustServista API has this capability and can take into account factors
such as named entities, context (who/what/where), clickbait title, and sentiment, however
it is acknowledged that there has not been much academic research conducted evaluating
this API due to lack of transparency [69].

13
2.3.1 Concepts Behind Natural Language Processing Dimensions

Considering which communication dimensions, to be translated into NLP dimensions, are

most important for identifying what can be biased or opinionated news, a total of 13 dimen-
sions stand out that can be identifiable using specified word embedding vocabularies. While
not all of these dimensions are present in every news article, these dimensions should be
universally considered when evaluating the news article for misinformation detection based
on this literature review.

1. Sentiment: Sentiment is potentially the most characteristic dimension of subjective

news [43], and it emotionally impacts readers to believe in false messages in the article.

2. Persuasion: This dimension is the same as bias and opinion; language that attempts
to get the reader to align with the article’s point of view is a very likely indicator of
misinformation [26].

3. Exaggeration: Exaggerated and/or outraged language is also indicative of misinfor-

mation [5].

4. Context: The more context that there is with cited persons of interest, the less likely
that there is misinformation present in the article [3].

5. Inclusion of multiple perspectives: Multiple perspectives that take in different

points of view from people of interest add more layers of context to the article and
reduce bias [3].

6. Named entities: Named entities can be very important for adding more context to
a news article and decreasing the likelihood of misinformation, especially if there is a
greater diversity of named entities [16]. The less diversity in named entities present in
an article, the more likely it is for misinformation to be present in the article.

14
 7. Use of statistics: Statistics can be very effective tools for clarifying facts in a sci-
entific manner and can add more context to a news article’s message. Even mentions
of dates and monetary figures can fall under the umbrella of statistics as these can
clarify exact timeframes and figures. More statistics means more likelihood that a
news article is trustworthy. This dimension is not commonly thought of when it comes
to disinformation detection, however it is important enough to provide well-rounded
context to an article [3].

8. Referencing of previous articles: As an alternative to cross-referencing, if the

article references certain information from a previous news article or interview, or
evidence from a previous report, that can add to the reliability of a news article [82].

9. Term frequency: If there are terms that are irregularly compared to words that are
normally used, this can lead to more likelihood of misinformation [58].

10. Distraction: Words that indicate an attempt to distract or deflect from the main
topic of the article can lead to a greater likelihood of misinformation [40].

11. Verification of claims: As an alternative to fact-checking, this dimension can provide

factual clarity to claims that are made in a news article and add to the trustworthiness
of the article [82].

12. Logical coherence: Words such as "first", "then", "therefore", and "consequentially"
indicate a logical flow of ideas and subjects in a news article, which can add more
context [82].

13. Clickbait title: As it is the only dimension related to the article title rather than the
article text, words present in the title that indicate clickbait or exaggeration can be
indicative of an article not being trustworthy [50].

15
2.4 Related Work

This section will cover a generalized overview of research in the data science field on misin-
formation detection. More related work regarding the specific components of the thesis will
be covered later in this chapter.

With the rise of misinformation as an issue to be tackled, it has also risen as an issue of
popularity within the scientific community, especially in the field of data science. In previous
field research, it has generally been established that supervised learning is required to train
a model to predict misinformation, and simple machine learning models including Naive-
Bayes classifier, support vector machine, random forest, and XGBoost have been used and
compared in their accuracies for classifying misinformation [57]. As the transformer model
is the most advanced model today that can predict misinformation [74], before its introduc-
tion deep learning models such as neural networks were the most efficient machine learning
models available for this task. For example, a Long Short Term Memory Recurrent Neural
Network (LSTM-RNN) was used for classifying misinformation on a binary scale (i.e., 0 for
misinformation, 1 for not misinformation) [14]. Deep learning discriminative, generative, and
hybrid models including recurrent neural networks (RNNs), deep belief networks, and con-
volutional restricted Boltzmann machines have been used and surveyed for misinformation
detection [29]. Text pre-processing mechanisms including regular expressions, lemmatiza-
tion, stop word removal, conversion to N-gram vectors and term frequency-inverse document
frequency (TF-IDF) sequence vectors have been used for converting text into vectorized
inputs for deep learning models imported from the Python Keras package for fake news de-
tection [38].

Not only various models have been used for misinformation detection, but also different
types of data have been utilized in previous field research. From politics to health, social
media post format to news article format, many categories of data have been used for mis-

16
information detection. Machine learning algorithms have been used to filter out opinion
spam, or comments that are repeatedly posted to push an opinion or viewpoint [49]. The
topic of COVID-19 has become a recent catalyst for misinformation research, for example
on YouTube videos as data containing COVID-19 misinformation [60]. Classification data
has even included COVID-19 misinformation related to cannabidiol products and warning
letters from the Centers for Disease Control and Prevention advising against using cannabis
products to cure COVID-19 [71].

Research has also been conducted on misinformation detection data that is not in the En-
glish language. Misinformation is not only a problem in the U.S., but it is also a problem
with a global impact. Misinformation that is not in English can be harder to detect due to
the presence of the foreign-language effect obscuring intuition when it comes to determining
misinformation potential [72].

2.5 Large Language Models

2.5.1 Theory Behind Supervised Learning and Transformer Model

Supervised learning is a machine learning approach that is designed to train a model to

correctly predict output values given specified input values. This is usually by training, or
making the model learn how to predict correctly, by using a pre-labeled training dataset
[57]. Supervised learning models have existed for decades for machine learning purposes,
including logistic regression, Naive Bayes, support vector machines, random forests, and
decision trees, and these can be used for classification and regression tasks [8]. Specifically,
detecting misinformation has been commonly used as a supervised learning task, with an
example being training a model to classify whether a news article is true or fake [57].

The large language model (also known as the transformer model), which is a more advanced
supervised learning model, was only recently invented in 2017 by Google [74]. Figure 2.1

17
shows the general architecture of the transformer model. The transformer architecture is
divided into an encoder and a decoder. The encoder consists of self-head attention layers
and feed forward network layers that accept input sequences. Using these transformed inputs
from the encoder, multi-head attention layers (and additional feed forward network layers)
in the decoder help capture the relationships among the positional embeddings in the inputs,
and then generate the outputs.

Figure 2.1: Transformer Model [74]

18
It is very crucial to note that transformer models in their inherent training nature are
not necessarily supervised. The original transformer model developed by Vaswani et al.
(2017) was trained in a semi-supervised setting, meaning that it was trained on a small
amount of labeled data but used the rest of the data that was unlabeled to make predictions.
Similar methodologies were made for training the LLMs that will be discussed in this section.
However, in the context of this thesis, the LLMs will be regarded as supervised learning
models; although the models are pre-trained in a non-supervised setting, fine-tuning the
LLMs to perform NLP downstream tasks will require labeled data and a supervised learning
approach [25], [42].

2.5.2 Bidirectional Encoder Representations from Transformers

Bidirectional Encoder Representations from Transformers, also known as BERT, was first
proposed by Google in 2018 [18]. This model was pre-trained using unlabeled data in an un-
supervised setting to perform NLP tasks including masked LM and next sentence prediction
[18]. It also incorporated significant improvements upon previous language representation
models, which had flaws of unidirectional NLP tasks and left-to-right context evaluation
[18]. BERT’s ability to fuse the left and right contexts of sentences together to evaluate text
holistically, and its versatility in its ability to do text prediction, question answering, and
text classification [18] made it a powerful tool for various NLP tasks.

19
 Figure 2.2: BERT Model Architecture (Pre-Trained) [18]

Figure 2.2 shows the architecture of the pre-trained BERT model. Since BERT is trained
to do next sentence prediction (NSP) and masked language modeling (MLM), these involve
randomly masking parts of the input sequences to train the models to guess the next sentence
or hidden tokens. Using the original transformer architecture from Vaswani et al. (2017),
the input embeddings that are to be put into the BERT model are the sum of the token
embeddings, the segmentation embeddings, and the position embeddings [18].

2.5.3 Robustly Optimized BERT Approach

Robustly Optimized BERT Approach, also known as RoBERTa, was proposed by Facebook
AI Research in 2019 as an improvement upon the original BERT model [42]. The model was
pre-trained using unlabeled data but fine-tuned using labeled data [42], which is also known
as self-supervised learning. Liu et al. (2019) argued that BERT had certain performance
training flaws that could be improved; the upgrades that RoBERTa incorporated included
training the model longer, using larger batches and sequences of textual data, and dynam-

20
ically changing the masking pattern applied to the training data [42]. Incorporating these
changes showed that the accuracy and performance of NLP downstream tasks on various
open-source datasets improved significantly using RoBERTa compared to BERT; the longer
the model was trained, the more these results reflected that performance [42].

2.5.4 Distilled BERT

A distilled version of the original BERT model, known as DistilBERT, was proposed by
Sanh et al. (2019). Since BERT was trained on a larger number of parameters and required
a large amount of data to be trained on, it was computationally expensive to train and
perform NLP tasks, therefore DistilBERT was introduced as a computationally faster and
lightweight alternative [59]. DistilBERT is 40% smaller than BERT and retains 97% of
BERT’s performance; its accuracy in performing downstream NLP tasks is similar to BERT
[59]. As seen in Figure 2.3, the architecture functions by incorporating less transformer
layers, containing the same multi-head attention and feed forward mechanisms, than BERT-
base’s original layers while preserving BERT-base’s efficacy.

Figure 2.3: DistilBERT Model Architecture [1]

21
2.5.5 Applications of Large Language Models to Natural Language

Processing Tasks

BERT, RoBERTa, DistilBERT, and many other LLMs are available for open-source use on
the Hugging Face website https://huggingface.co [80]. These models can be downloaded
using the Transformers package in Python and fine-tuned to perform any NLP task including
chat generation, sentence prediction, text classification, and text summarization. Fine-tuned
models designed to perform specific NLP tasks can also be deployed to the website. Generally,
the maximum number of input tokens that these models can process is 512 [18], which
is equivalent to around 400 words. OpenAI’s ChatGPT, although regarded as AI, is at
its foundation an LLM as it has the ability of being a chatbot with multiple NLP-based
applications including assisting with writing papers, assisting with programming code, and
being an interactive learning tool by providing answers on any topic [45]. In addition, LLMs
have been used for misinformation classification tasks on textual data, and these transformer
models have overall provided good results for performing these tasks [13], [33], [54]. Finally,
comparing multiple LLMs including BERT, RoBERTa, and DistilBERT to see which model
performs most effectively in NLP tasks such as misinformation classification has been utilized,
as model performance comparison is a common practice in the data science field [42], [59],
[57].

2.5.6 Applications of Large Language Models to Application

Programming Interfaces

APIs have likewise made an impact on AI accessibility [51]. ChatGPT currently has a pay-
as-you-go API with different models depending on the intended maximum number of tokens
used. As ChatGPT is also a large language model, the ChatGPT API serves as an example
integration of LLMs and APIs to not only provide multi-dimensional information from the
API’s website server, but also actual model predictions from the LLM itself.

22
2.6 Unsupervised Learning

2.6.1 Theory Behind Unsupervised Learning

Unsupervised learning is a different type of approach in machine learning that sharply con-
trasts with the approach of supervised learning. In supervised learning, the model is trained
based on pre-labeled data, however in unsupervised learning the model utilizes unlabeled
data, allowing the model to generate patterns and insights among the data. Common un-
supervised learning tasks include clustering of data points and multi-dimensional feature
selection [35].

2.6.2 Association Rule Mining

Association rule mining is generally performed on transactional data to determine which

items in a transaction go together, and this method can be found in open-source Python
packages including Apriori or FP-Growth [39]. Two common examples of association rule
mining are determining which items in a grocery shopping list are generally bought together
and determining which movie or TV show choices on a streaming service are watched to-
gether. Data preprocessing for this method is usually done by generating frequent itemsets
among the data through one-hot encoding the terms that are present or not present in the
transaction (i.e., assign a value of 0 if not present and assign a value of 1 if present). Then
given these frequent itemsets, association rules are generated based on how commonly the
items in the itemsets appear and which items are most closely associated with one another in
the transactional data. Figure 2.4 shows an example of general processing with association
rule mining.1
1
Reference link for the image in Figure 2.4: http://www.big-data.tips/association-rules

23
 Figure 2.4: Association Rule Mining Example

2.6.3 Anomaly Detection

Anomaly detection is the unsupervised learning technique of finding patterns in data that
do not conform to normal behavior [9]. Data preprocessing is done by converting data
into a vectorized matrix using TF-IDF, bag-of-words, or other word embedding techniques.
After defining the number of reduced features for analysis and the percentage of intended
outliers, or anomalies, in the data, an isolation forest model is then initialized and trained to
determine which data points are considered anomalies in the data; an isolation forest is an
ensemble of isolation trees that are built for a dataset, and the anomalies that are determined
by the isolation forest are the observations that have the shortest average path length on the
isolation trees [41]. For example, setting the model with a contamination of 0.05 means that
5% of the data points will be designated as anomalies based on their reduced feature values,
and data visualizations illustrating anomaly detection can show some data points that are
colored differently compared to the majority of the data points.

24
2.6.4 Cosine Similarity

Cosine similarity is a text similarity metric that is commonly used for text analysis. This
metric models a text document as a vector of terms, and the similarity between any two
documents can be derived by calculating the cosine value between the two documents’ term
vectors. Cosine similarity can have applications to semantic similarity, or how similar texts
are regarding contextual meaning [53]. The formula below represents the cosine similarity
between two text vectors A and B, where cos(θ) is the cosine of the angle between the two
text vectors, and ∥A∥ · ∥B∥ are the Euclidean norms of the vectors A and B.

Pn
A·B Ai · Bi
cos(θ) = = pPn i=12 pPn
∥A∥ · ∥B∥ i=1 Ai ·
2
i=1 Bi

2.6.5 t-Distributed Stochastic Neighbor Embedding

Stochastic neighbor embedding is a dimensionality reduction approach that is used to calcu-

late the Euclidean distances between data points as reduced feature similarities [73]. t-SNE
is an expansion of that, as it provides a solution to visualizing high-dimensional data in a 2-
or 3-dimensional space, resulting in the generation of differentiated clusters of data points
based on their reduced feature values [73]. An example of such visualization, data visualiza-
tions showing results of t-SNE analysis can display differently colored or labeled clusters of
data points in a 2- or 3- dimensional scatter plot [73].

2.6.6 Latent Dirichlet Allocation Modeling

LDA modeling, also known as topic modeling, is a common unsupervised method used for
generating topics based on an input corpus of text documents [32]. The LDA model can be
found for open-source use in the Gensim Python package. The LDA model assumes that
the corpus contains texts over a variety of topics and that each topic contains a certain
distribution of words. The model iterates through each of the documents until a certain

25
number of topics with keywords are generated, and subsequently the topics can be inferred
based on the keywords found by the model.

Figure 2.5: LDA Model [61]

2.6.7 Applications of Unsupervised Learning to

Misinformation Detection

Some of the aforementioned and other unsupervised learning methods have been used in
misinformation detection [27], however research using these models has not been as extensive
compared to research using supervised learning models. Association rule mining has been
used for determining patterns among misinformation in women’s health tweets [37]. Various
anomaly detection models have been used for classifying financial misinformation [64]. t-SNE
clustering was used to plot the topic clusters of texts to show visual relationship patterns
between topics in misinformation data [14]. LDA modeling in conjunction with calculating
cosine similarity scores has been implemented in misinformation data training and has been
shown to produce accurate results [46]. Topic modeling has also been shown to produce
topics that are likely to be associated with misinformation claims [81]. Given these studies,
using these unsupervised learning methods to determine hidden data patterns between true

26
and fake news articles could potentially be of great value.

2.7 Generative Artificial Intelligence Models

2.7.1 Theory Behind Artificial Intelligence

AI is the concept of intelligent machines or computers designed to emulate human intelligence

[36]. As AI is based on neuroscience and how the human brain functions, neurons were able to
be mathematically implemented as early as 1943 [44]. This breakthrough led to the eventual
development of the artificial neural network (ANN) in the 1980s, which is currently the most
widely used model in AI research [19]. Prior to the rise of AI applications like ChatGPT
in late 2022, advancements in supercomputing power and big data technologies led to the
acceleration of AI development, thereby making effective cognitive computing easier [19].
Deep (machine) learning algorithms like ANNs can also be credited for the renewed success
of AI research endeavors before the popular emergence of ChatGPT [19].

Machine learning and neural networks reflect a cognitive science application of AI, while
NLP is part of a natural language interface application of AI, therefore these subjects are
not the same as, but are rather subareas of, the general field of AI [36]. Over the years, AI
has evolved to be implemented through LLMs and NLP [56]. This has led to the emergence
of generative AI, a class of AI language models that can create new data based on patterns
and structures learned from existing textual data; ChatGPT is an example of generative AI
[56]. The transformer model architecture [74] has likewise been a staple foundation of AI
models; this type of model helped introduce solutions to limitations of previous sequence-
to-sequence models like the convolutional neural network (CNN) and the recurrent neural
network (RNN) [56]. The recent advancements in generative AI, including having a language
model as its base, have led to AI gaining much popularity through its superior performance
as a generative chatbot and easy accessibility [56].

27
2.7.2 OpenAI ChatGPT

OpenAI’s ChatGPT was released to the public in November 2022 as a user interface chatbot,
using its version 3.5 as the model architecture [56]. ChatGPT was trained on a large corpus
of text data and fine-tuned on a specific task of generating conversational responses, which
allows it to generate human-like responses to user queries [56]. ChatGPT also has an API
where the user can specify the version and number of input tokens desired for data collection.
ChatGPT is known to have some limitations including being trained on data that is only as
recent as 2021 and being trained on data that can contain biased language [17], [56]. Other
significant concerns about using ChatGPT include data privacy, production of factually
inaccurate or unreliable responses, and inability to capture the contextual understanding of
prompts [56].

2.7.3 Google PaLM

Google’s Bard was released in March 2023 and is a user interface chatbot similar to Chat-
GPT. Bard differs from ChatGPT such that the data that Bard is linked to is from Google’s
expansive engine search network, while training data for ChatGPT was not necessarily linked
to the internet [77]. The language model foundation for Google Bard is the Pathways Lan-
guage Model, also known as PaLM [77]. While Bard is Google’s user interface version of its
AI, PaLM 2 (released May 2023) is Google’s API version of Bard.

2.7.4 Applications of Using AI Models for Misinformation

Detection

GPT-3, a previous version of the current GPT model, has been observed as being a potential
spreader of misinformation [67]. Some AI models and social media networks have APIs for
easy data collection, however these can have the potential to spread disinformation [48]. In
a preliminary study, 72% of verified claims were correctly identified by ChatGPT [28]. This

28
statistic is modestly good, however it is possible that this number is not higher due to some
of the claims originating after the year 2021, where ChatGPT is most recently trained on
data pertaining to that period. A study involving evaluating ChatGPT-3.5, ChatGPT-4,
Bard, and Microsoft’s Bing AI showed that these had a moderately promising ability to
classify misinformation accurately, with ChatGPT-4 having the best accuracy with 71% and
the average overall accuracy among the AI models being 65% [7].

2.8 Research Gaps

After evaluating the literature review that has been conducted relating to this thesis, there
are some research gaps that this thesis can address. First, since most misinformation classifi-
cation tasks using LLMs have used binary labels [14], utilizing a regression-based framework
instead would allow the base model of the proposed tool to assign scores on a comprehensive
scale from 0 to 100, using the NLP dimensions of trustworthiness for more sophisticated
analysis. The NLP dimensions can also help address a gap regarding the Hugging Face
LLMs, which can only input a maximum of around 400 words. Thus, holistic evaluation of a
news article being greater than 400 words is difficult using only an LLM; there must be other
features, taking into account the entire context of the article text, to help improve the pre-
dictability of the trustworthiness scores. Explainability in misinformation classification tools
is another important research gap to address. AI models can have the power to spread and
counter misinformation, but it is important to promote greater transparency and "human-
in-the-loop" elements in future research [34]. Including explainability in the proposed tool
to justify which NLP dimensions were present or not present in an input text would make
a transparent and trustworthy addition, as some LLM-based AI systems currently lack ex-
plainability; this can lead to users not being able to understand why the model is making
the decisions that it is making and consequently, trust in the technology can be diminished
[56]. A research gap that was also discovered was that regardless of the development of

29
pre-existing tools used for misinformation detection, the data that these tools are trained on
will have to be periodically updated to include more up-to-date information [24], [63].

Some research has been done with regard to misinformation characteristics through associ-
ation rule mining, cosine similarity, t-SNE, and anomaly detection, but this research with
such unsupervised learning methods is not as extensive as research using supervised learn-
ing methods. Although a separate dataset of true and false statements could be used for
the unsupervised learning approaches will likely be limited to one topic with a sufficient
amount of data, these unsupervised learning approaches can be used on other pre-labeled
topic datasets, including FEVER and LIAR [68], [78]. However, the data that will include
some news articles will likewise be valuable to determine which invisible characteristics could
distinguish which news is true and which is fake. Such characteristics could include a re-
duced dimension value, a topic keyword, a distribution having a center significantly different
from another distribution’s center, a cosine similarity value, or a cluster of data points that
are significantly different from another cluster meaning a greater or lesser likelihood of mis-
information for a given news article. Data visualizations will likewise be instrumental in
illustrating these revelations in distinguishing the true and fake news. Some research on
such data visualizations (not pertaining to misleading data visualizations or images in news)
has been done [65], [82], however there is still a gap when it comes to visualizing vectorized
or textual differences between real and fake news.

Finally, ChatGPT and PaLM are recent AI models that currently appear to have the po-
tential to detect misinformation if prompted with an input text. Despite some research
having been done on their ability to detect misinformation [28], [7], more studies need to be
conducted to determine whether these AI tools can be trusted for misinformation detection
tasks and which advantages and disadvantages these tools have when detecting misinfor-
mation. These aspects can pertain to the classification accuracy as well as the perceived
explainability of these AI models.

30
2.9 Summary

The discussion of this literature review helped integrate communication dimensions of misin-
formation with the field of data science to open the possibility of creating data tools that can
identify misinformation potential. There are also the explorations of unsupervised learning
and evaluating AI models that have had some but not sufficient research done on their capa-
bility to effectively identify misinformation. This thesis will help address significant research
gaps connected to these areas in the domain of automated misinformation detection.

Figure 2.6: Tabular Summary of Literature Review

31
Chapter 3

Supervised Learning Using Large Language

Models and Natural Language Processing
Dimensions

3.1 Overview

The first type of model that will be analyzed in this thesis is the fine-tuned supervised
learning LLM. Pre-trained LLMs are usually trained in a self-supervised manner, however
fine-tuned LLMs can be trained in a supervised manner to perform downstream NLP tasks.
Supervised learning models are trained to make predictions, and LLMs are complex, versatile
NLP models that can execute analysis using the forward and backward context of input texts.
LLMs have been shown to produce promising misinformation detection results, however when
compared to the approaches of unsupervised learning and AI, is supervised learning using
LLMs necessarily the most effective approach? Or do other approaches reveal features that
can predict misinformation more accurately? That is what this thesis will attempt to resolve.

Thousands of news article texts covering a variety of topics will be used for training the LLM,
as it is important that the LLM be trained on as large and diverse a dataset as possible. The
Hugging Face LLMs BERT, RoBERTa, and DistilBERT will be trained on the dataset and
compared to determine which model has the better performance in predicting trustworthiness
scores. The models will also be trained regression-wise to produce trustworthiness scores

32
between 0 and 100 (0 meaning low trustworthiness and high misinformation potential and 100
meaning high trustworthiness and low misinformation potential) based on communication
dimensions of misinformation that are easy to embed as NLP dimensions; these dimensions
are easier to identify as features of biased or opinionated news. These NLP dimensions will
serve as quantified features that will help make the trustworthiness score predictions more
accurate.

LLMs usually take a maximum input of 512 tokens, and a significant portion of these news
articles are longer than 512 tokens, thus the model has to truncate the input texts and
not incorporate the entire context of the articles. By themselves, these truncated texts
will not help the LLMs predict the trustworthiness scores accurately, therefore the NLP
dimensions will be needed for capturing the entire context of the articles and fed into the
LLMs. Therefore, it is important that the chosen base model for the data science tool be
able to evaluate entire news articles, regardless of length, to bypass the limitations of the
LLMs to make accurate misinformation potential predictions.

3.2 Natural Language Processing Dimensions

13 NLP dimensions were chosen for creating the trustworthiness scores. The minimum
cumulative trustworthiness score that can be given by the model is 0, and the maximum
cumulative trustworthiness score that can be given is 100, as the separate dimension scores
will be summed together. Below is why these dimensions were chosen, which textual features
are indicative of these dimensions, and how much they will be weighted. All dimensions
except for referencing of previous articles, distraction, verification of claims, and clickbait
title were weighted a 10 as these dimensions appeared to have the most relevancy and impact
on trustworthiness, which will be discussed further in Section 3.5.4.

1. Sentiment: Are there words that indicate praise, fear, anger, etc. for the person or
topic of interest? Sentiment is an important indicator of biased or opinionated news,

33
 and due to its impact on whether an article is trustworthy or not, it will be given a
maximum weight of 10.

2. Persuasion: Are there words indicating biased rhetoric or attempting to coax the
reader to accept the article’s point of view? Also known as a dimension that is in-
dicative of bias or opinion, language that attempts to get the reader to align with the
article’s point of view is a very likely indicator of misinformation. This dimension will
be given a maximum weight of 10.

3. Exaggeration: Are there words or punctuation that indicate exaggeration or outrage?

Exaggerated and/or offensive language is a very definitive indicator of misinformation
and untrustworthiness of a news article, and its impact means that it will be assigned
a maximum weight of 10.

4. Context: If citing potential misinformation, is the person of interest saying it or is the

article saying it? Quotation marks are indicators that the article is directly citing a
person of interest, and context that includes whether the article is saying a statement or
if the person of interest is saying a statement is important in determining whether the
article is providing a platform for misinformation in its viewpoint. If the article provides
a biased viewpoint and does not objectively cite a person’s misinformed statement, then
that can be indicative of misinformation. This dimension will be assigned a maximum
weight of 10.

5. Inclusion of multiple perspectives: Are there words that indicate that someone
said, suggested, explained, etc. in the article? Multiple people being cited in the article
that have different viewpoints and defenses make the article more objective, less biased,
and with more context. This dimension will have a maximum weight of 10.

6. Named entities: Are there multiple named entities? A wide diversity and usage of
named entities, including people, organizations, dates, and locations are important for

34
 providing additional context to the article. This dimension has a large impact and will
be given a maximum weight of 10.

7. Use of statistics: Are there numbers or figures that are present in the article? Usage
of statistics provides even more context to a news article, but more so from a numerical
or mathematical standpoint. Polls, surveys, data trends, monetary figures, and other
numerical points can contribute a well-rounded, factual context, and can have a great
impact on determining whether a news article is trustworthy or not. This dimension
will be given a maximum weight of 10.

8. Referencing of previous articles: Are there words that indicate previous stories,
editorials, or op-eds? Some articles can cite previous reports, op-eds, and articles that
can give some context to the subject matter of the news article. However, since this
dimension may not be present in all news articles, this dimension will not have much
impact and will be assigned a maximum weight of 2.

9. Term frequency: Are there unique terms or words that compare to the rest of the
corpus of news articles? Unique terms that may be different from normal words used
in a news corpus can be indicative of misinformation. These uncommon terms may
raise concern about how biased or opinionated the news article is. This dimension will
have a maximum weight of 10.

10. Distraction: Are there words that indicate distraction or contradiction from the main
topic of the article? Some articles may attempt to distract the reader by using certain
words that divert from a given topic of interest that may be important to address.
However, this tactic is not used in all articles, therefore this will have a maximum
weight of 2.

11. Verification of claims: Are there words that acknowledge that a claim has been
verified, alleged, debunked, etc.? While fact-checking objective statements is difficult,

35
 there can be ways to quantify words that acknowledge that certain claims were alleged,
confirmed, debunked, etc. such that it can add an alternative. However, since some
articles do not have this, it will be given a maximum weight of 4.

12. Logical coherence: Are there words that indicate logical sense or flow? A logical flow
of the article or viewpoint can provide logical context and can also indicate whether
the text is well-written. This dimension will have a maximum weight of 10.

13. Clickbait title: Are there words in the article title that indicate potential clickbait?
A clickbait title can be indicative of an article that may likely not be trustworthy. This
dimension will have a maximum weight of 2.

Figure 3.1: Example of Article Text with Low Trustworthiness Score

36
Figure 3.1 shows an example of an article text from the training dataset that had a low
trustworthiness score. Note that this is not the entire text as this article was very long and
had to be truncated for this example image. Some words have been censored in the figure
due to offensive language. The highlighted portions show specific textual features that likely
affected its trustworthiness score based on the dimension vocabulary matches in the text,
thereby the reasons why this article text was given a low trustworthiness score include:

• Very high use of exaggerated language.

• High use of persuasive language and some sentiment.

• No use of statistics.

• No referencing of previous articles, verification of claims, or logical coherence.

• No context or inclusion of multiple perspectives.

• Some use of named entities.

• A positive is that the article title is not clickbait and is objective.

37
 Figure 3.2: Example of Article Text with High Trustworthiness Score

Figure 3.2 shows an example of an article text from the training dataset that had a high
trustworthiness score. Note that this is not the entire text as this article was very long and
had to be truncated for this example image. The highlighted portions show specific textual
features that likely affected its trustworthiness score based on the dimension vocabulary
matches in the text, thereby the reasons why this article text was given a high trustworthiness
score include:

• No use of sentiment nor exaggerated language.

• Moderate use of statistics and numerical figures.

• High use of named entities, context, and multiple perspectives.

• No clickbait title.

38
 • High use of verification of claims.

• A negative is there is little to no logical coherence.

3.3 Data Collection

The news article dataset was compiled using various datasets and subsets extracted from
Kaggle and GitHub1 [47], [62]. Preferred facets of these separate datasets to make compi-
lation easier included the dataset having the fields of article origin (i.e., publication, URL,
domain), article title, and article text, and that the dataset was in a .csv file for reading in
by the Pandas dataframe package in Python. The article title and article text were the most
important fields as these contained the necessary NLP dimensions to extract for analysis and
prediction. The article origin was used for reference purposes, however when creating the
proposed data tool the ultimate objective is for the tool to detect misinformation requiring
as few fields as possible and to make predictions based only on the text and not on external
factors of misinformation that include article origin, article author, etc. If the tool required
additional fields, it should consequently be considered that potential users of the tool may
not have or may have difficulty retrieving that information, thereby leading to incomplete
or inaccurate predictions from the model.

These are the following sources from where the data originate:

• Tinker Air Force Base Supply Chain Risk Management project dataset

This data was extracted from work on the TAFB SCRM project with the Data Institute for
Societal Challenges (DISC). It contains webscraped news articles and blog posts.

• https://www.kaggle.com/datasets/snapcrack/all-the-news/data
1
See also the following URL sources:
https://www.kaggle.com/datasets/snapcrack/all-the-news/data,
https://github.com/several27/FakeNewsCorpus,
https://github.com/pmacinec/fake-news-datasets

39
This data was from the "All the news" Kaggle dataset and contains news articles from a
variety of mainstream and independent news sources. Figure 3.3 shows a bar graph depicting
the counts of news articles from the article publication origins featured in the dataset.

Figure 3.3: Bar Graph of "All the news" Publications

• https://github.com/qwerfdsaplking/MC-Fake

This data was from a GitHub dataset that contained news articles from a variety of topics
(politics, entertainment, health, COVID-19, and the war in Syria) [47].

• https://github.com/several27/FakeNewsCorpus

40
This data was from a GitHub dataset that contained mostly independent news articles that
were categorized as satire, extreme bias, conspiracy theory, hate news, etc. However, only
the 250-size open-source sample was used as the entire dataset required an unarchiver app
to read the actual larger version of the data.

• https://github.com/pmacinec/fake-news-datasets

Multiple datasets can be found in this GitHub repository, however the ones that were chosen
were called "fake_news_detection_kaggle" and "getting_real_about_fake_news_kaggle",
and these datasets originated from Kaggle. The first dataset listed contains a variety of news
articles that contain both real and fake news and the second dataset contains fake news
articles that were extracted using the webhose.io API.

• https://github.com/Gautamshahi/FakeCovid

This data contains news articles related to COVID-19, a popular topic for fake news detection
modeling [62].

3.4 Data Preprocessing

Although the automatic provision of the fields of article origin, article title, and article text
in the .csv datasets did assist greatly in preparing the training dataset, many steps had to
be taken to preprocess the data.

The separate .csv datasets were loaded into a Jupyter notebook to make it easily accessible to
view the various stages of the datasets as they were preprocessed. Even though JupyterLab
was initially used for programming during the course of this thesis, Google Colab ended up
being used for the majority of coding as this utilized RAM and GPU resources from the cloud
rather than resources from the local machine. The following steps were then taken on each
dataset before the cleaned dataset was exported to a .csv. First, if the dataset was very large
(50,000 articles were needed in total), then random subsets were taken of those datasets to

41
ensure consistency of the data. Any observations that had a null article origin, article title,
or article text were removed entirely as having a nonexistent field would cause issues for the
model to make predictions. Unlike preprocessing numerical data where null values can be
imputed using the field distribution’s mean to retain information, the loss of any fields when
it comes to textual data automatically signifies that entire observation cannot be used as it is
impossible to impute text. News articles that were not in the English language were removed
as well; since the NLP dimensions will be embedded using certain vocabularies in English,
vocabularies in foreign languages would require a lot of time and difficulty to compile, and
could require the cost of a translation API. Note that this would impose the limitation that
the proposed tool would not be functional using foreign language news articles as input.
Most texts were above 700 characters in length to include as much context as possible for
analysis, however there were some texts under 700 characters that were kept. Nonetheless,
an objective of this tool is to make predictions on any news article text regardless of length.
Some filler texts were removed from the dataset as these were texts that contained generic
information, tended to duplicate each other, and did not provide sufficiently lengthy nor
diverse context for analysis.

Finally, as these preprocessing techniques were used on each of the separated datasets and
exported as cleaned datasets to .csv files, these cleaned files were then imported back into
a Google Colab notebook and merged together using the same column field names "Article
Origin", "Article Title", and "Article Text". This joining created a training dataset of
50,092 news articles. In addition, \n characters were removed from the article texts as many
of these texts contained these characters, which would be unnecessary for analysis using
LLMs. However, most of the texts did not undergo further preprocessing as since LLMs
take as much context as possible for analysis and do not require common text preprocessing
methods such as lemmatization and stopword removal, it was important to retain as much
of the context of the news articles as possible for accurate predictions from the LLMs.

42
3.5 Exploratory Data Analysis

Looking at the finalized starting dataset, exploratory data analysis was performed on the
fields in the dataset and this initial analysis provided some useful insights into the data that
will be heavily analyzed during the course of this thesis. There was also exploratory analysis
conducted on the NLP dimensions, which helped determine how much each dimension would
be weighted for the trustworthiness scoring.

3.5.1 Article Text

First, the cleaned article texts were examined based on length (in characters and words)
to determine the approximate distribution of the text length. Statistical summaries and
histograms with kernel density curves were made to illustrate the distributions.

(a) Article Text Length (Characters) (b) Article Text Length (Words)

Figure 3.4: Statistical Summaries

The statistical summaries show that the median character length of a news article in the
training dataset is 2,946 characters. In words, that is equivalent to 485 words. The mean
word length of an article text is approximately 668 words, and 25% of all articles in the
training dataset are 826 words or above in length.

43
 (a) Article Text Length (Characters) (b) Article Text Length (Words)

Figure 3.5: Histograms with Kernel Density Curves

Figure 3.5 shows the distributions of the article text lengths, in characters and words, as
being heavily skewed right. This is potentially due to not many news articles that exist that
are thousands of words in length (the tail in the word length distribution seems to disappear
at around 5000 words). Despite the distributions being skewed right as if the articles appear
to be short in length, the statistical summaries show that there is indeed a large portion of
articles that are rather lengthy.

Therefore, since at least half of the articles in the training dataset are above the maximum
input length of 512 tokens, or about 400 words, for a Hugging Face LLM, there is a more
emphasized need for including other features in the proposed data tool that can evaluate the
article texts in their entirety, going beyond the limitation of the restricted number of input
tokens for an LLM.

3.5.2 Article Title

Next, the article title field was analyzed. Analysis was emphasized on word length rather
than character length as unlike the article texts the article titles were not filtered out based
on a certain character length. Thus, the length of the article titles did not matter as long as
the text was not null in the dataset.

44
 Figure 3.6: Article Title Length Statistical Summary

Figure 3.7: Article Title Length Histogram

Figures 3.6 and 3.7 show that the minimum number of words in an article title was 1
(therefore no null or blank values in the article title field per the intended preprocessing),
and very few article titles had fewer than 4 words. The mean and median number of words in
an article title were around 10, and the maximum number of words was 158. The distribution
of the article title length was still skewed right, however not as skewed as the distribution
of the article text length. Based on the respective distributions of article text length and
article title length, the overall average of an article title length is very short compared to

45
the article text length. Therefore, there should still be more emphasis on analysis of the
comprehensive article text, in addition to the small but substantial role that the article title
plays in misinformation detection.

3.5.3 Article Origin

The final field in the dataset that was explored was the article origin field. Since the article
origin field was a mixture of article publication names, article URLs, and article domains, the
initial number of unique article origins was 27,486. This was primarily because each article
URL was unique and had no duplicates, whereas some article publications in the sample
format "Fox News" had multiple articles duplicating this publication origin. After simplifying
most of the article URLs and domains under singular publication origins to determine which
article origins had the highest counts of articles, there were 1,555 simplified unique article
origins. Note that some of these were still unique URLs from the same publication source,
however there were very few of these observations remaining.

Figure 3.8: Top 20 Article Origins by Count

46
 Figure 3.9: Top 10 Article Origins by Count

Figures 3.8 and 3.9 depict the top 20 and top 10 article origins by count, and the percentages
that are represented in the pie charts are compared to the rest of the top 20 or 10. The top
10 article origins comprised 19,817 articles, or 39.5% of the articles in the training dataset.
The top 20 article origins comprised 27,828 articles, or 55.5% of the articles in the training
dataset. As for the remainder of the dataset, there are still 1,535 unique article origins
that comprise the last 45.5% of the articles. Other notable article origins included Reddit,
Microsoft Blogs, Disclose.tv (a far-right publication based in Germany), Fox News, various
cybersecurity blogs, and various celebrity news outlets. The exploratory analysis of the
article origins has shown that besides mainstream media, there is a large diversity of article
origins in the dataset that is sufficiently suitable to train the LLMs to make predictions for
a wide variety of news articles.

47
3.5.4 Natural Language Processing Dimensions

Exploratory data analysis was done on the 13 NLP dimensions that were generated for the
articles in the dataset, and the distributions for these articles were analyzed to determine
which weights would be assigned to the dimensions to calculate the overall trustworthiness
scores for the articles.

(a) Sentiment (b) Persuasion

(c) Exaggeration (d) Context

Figure 3.10: NLP Dimension Distributions (1-4)

48
(a) Multiple Perspectives (b) Named Entities

(c) Statistics (d) Referencing of Previous Articles

(e) Term Frequency (f) Distraction

Figure 3.11: NLP Dimension Distributions (5-10)

49
(a) Verification of Claims (b) Logical Coherence

(c) Clickbait Title

Figure 3.12: NLP Dimension Distributions (11-13)

50
As seen in the figures above, some of the dimensions such as sentiment and exaggeration
do not appear to have much spread, however since these dimensions are the most indicative
of misinformation these were assigned weights of 10. The use of statistics dimension is
very skewed, however its center of spread is greater than the other dimensions’ distributions
(aside from named entities). Although the dimension of logical coherence appears to have
a smaller center of spread compared to other distributions, there was more variability in
the data, therefore it is a dimension that can have a significant impact on determining
misinformation. The dimensions of referencing previous articles, distraction, and verification
of claims did not appear to have as much spread nor a greater center of spread compared
to other distributions, therefore this is why they were weighted less (however, verification
of claims was still crucial enough to be weighted more since it serves as an alternative to
fact-checking). Finally, the dimension of clickbait title, based on its distribution spread,
appeared to have the least impact in determining misinformation, which is why it was given
a weight of 2. It could be considered in the future to increase the weight for this dimension
as it can plausibly be a significant indicator for biased news, however the initial exploratory
analysis shows insufficient evidence to support increasing the weight for that dimension. As
shown in Figure 3.12(c), the vast majority of articles do not have clickbait titles even though
some articles may be very biased in their texts, and increasing the weight of the clickbait
title dimension would require a more extensive vocabulary for quantifying this dimension.

3.6 Data Labeling

Manually labeling the 50,092 news articles in the training dataset based on the proposed 13
NLP dimensions would have been an extremely difficult and time-consuming process. Thus,
a more automated approach to labeling the articles had to be devised. A vocabulary list was
made for each NLP dimension (see Chapter 10), with the exceptions of the named entities
and term frequency dimensions. These vocabulary lists contained words and punctuation

51
that would generally be indicative of the dimensions. The vocabulary for the named entities
dimension would be extracted using the spaCy package in Python, and the term frequency
dimension would not require a vocabulary list, but rather the TF-IDF vectorizer in Python.
This would vectorize all of the words in a news article text based on term frequency in
relation to the entire corpus of news article texts.

Next, using the NLTK tokenizer in Python, the article texts would be iterated through and
the number of times that a token in the dimension vocabulary appeared in an article text
would be counted. Then the dimension presence count would be divided by the length of
the tokenized text to normalize the count based on the length of the article, quantifying the
NLP dimension. In Figure 3.11(b), the normalized count for the named entities dimension
can be over 1 because the length of the named entities list extracted from the article text
would be divided by the tokenized length.

Finally, taking the distributions of the normalized counts for each dimension, percentiles
were calculated to determine which thresholds in the distribution would mean a certain
number of points in the dimension weight to be allocated. For example, if the normalized
persuasion count was less than or equal to 0.004, the full 10 points would be given for the
dimension as this number indicates little to no persuasion present in the article text. As a
contrary example, if the normalized logical coherence count was greater than 0.0125, the full
10 points would be awarded since this threshold indicates that there is a high presence of
logical coherence in the article text. After the individual dimension scores were allocated,
these scores were summed up to calculate the cumulative trustworthiness score on a scale
from 0 to 100. Figures 3.13 and 3.14 show the distribution of the final score labels.

52
 Figure 3.13: Distribution of Score Labels

Figure 3.14: Statistical Summary of Score Labels

The distribution of the score labels follows a Gaussian, or normal, distribution, with a cen-
ter of spread (the mean) at around 52. 50% of the score labels are below 53, the minimum
automated score label is 4, and the maximum automated score label is 84.

Some data exploration was also done by dividing the labeled news articles into the categories
"Likely Trustworthy" and "Likely Untrustworthy". If the labeled cumulative trustworthiness

53
score was below a 50, it would be classified as "Likely Untrustworthy", and if the cumula-
tive score was above a 50, it would be classified as "Likely Trustworthy". This classification
threshold of 50 would likewise be utilized for classification tasks later on in the thesis. 18,964
articles were "Likely Untrustworthy" and 31,128 articles were "Likely Trustworthy". In ad-
dition, 17,435 articles out of the training dataset had a trustworthiness score between 45 and
55, the closest interval in the middle that could determine which articles would be trust-
worthy and which ones would not be. While many articles were close to the trustworthiness
score threshold of 50, these articles tended to have little to no context, logical coherence,
inclusion of multiple perspectives, nor use of statistics, meaning a large deduction from the
trustworthiness score that could determine, by a few points, whether an article could be
trustworthy or not.

3.7 Model Training

The following steps discussed in this section were taken for the data preparation and training
for all Hugging Face LLMs. First, the quantified NLP dimensions for the given article text
were concatenated with the text to improve the accuracy of the LLM. The dataset would
then be split up into 70% training, 15% validation, and 15% testing. All LLMs were imported
from the Transformers package in Python and were trained using the TensorFlow package.
Although the PyTorch package, which is also commonly used for training LLMs, was tested,
the models appeared to train slower and less accurately using PyTorch compared to using
TensorFlow.

The training, validation, and testing sets were tokenized using the imported LLM tokenizers
and converted into input IDs, or token indices that are numerical representations of the
token sequences that will be accepted by the LLM. The input IDs would then be compiled
into a Keras Input layer for the model. A Keras Dense layer creating the regression head for
the model was also created to fine-tune the model. After the overall model with the layers

54
was initialized, the model would be compiled with the Adam optimizer with a learning rate
of 1e−5 , mean squared error (MSE) as the loss function, and mean absolute error (MAE)
as an additional evaluation metric. The batch size for training all LLMs was 32, and Keras
early stopping was implemented to track and save the epoch where the model had the least
validation loss; the patience was set at 2 epochs until no improvement in decreasing the
validation loss was evident. The combination of the Adam optimizer, 1e−5 learning rate,
and batch size of 32 were consistently chosen to have fast and efficient training of the LLMs
while keeping training stable so as to not have the loss function converge too quickly. For
testing evaluation metrics, MSE, MAE, Pearson Correlation Coefficient, Spearman Rank
Correlation Coefficient, and R2 were used. All model weights were saved to be later imported
for making trustworthiness score predictions.

3.7.1 BERT

The imported tokenizer and pre-trained model from the Transformers package were "bert-
base-uncased". The maximum input token sequence length that was used by the model was
176; maximum input sequences higher than the designated token length led to GPU overload
in Google Colab, and this applied to all LLMs. In addition to input IDs, the BERT model
also used attention masks, or encodings that match the length of the input IDs of one text
to the input IDs of another text. Attention masks are an optional argument while training
LLMs, however they are recommended for controlling variable sequences. The input IDs
and attention masks along with the score labels were converted into TensorFlow datasets for
input into the BERT model.

55
 Figure 3.15: BERT Model Architecture Summary

3.7.2 RoBERTa

The imported tokenizer and pre-trained model from the Transformers package were "roberta-
base". The training of the RoBERTa model also utilized attention masks. The maximum
sequence length determined was 160 tokens. As the BERT model only used the regression
head for output, the RoBERTa model instead used the regression head with CLS (classifier)
tokens. Before this model was trained, the tokenized training, validation, and testing sets
had the input IDs and attention masks extracted and converted into NumPy arrays, and the
score labels were also converted into a NumPy array for training to function.

56
 Figure 3.16: RoBERTa Model Architecture Summary

3.7.3 DistilBERT

The imported tokenizer and pre-trained model from the Transformers package were "distilbert-
base-uncased". The determined maximum sequence length was 300 tokens, which can be
explained by DistilBERT being a much lighter version of BERT that can retain most of
BERT’s performance. No attention masks were used in the training of the DistilBERT
model, and the regression head contained a pooling output.

57
 Figure 3.17: DistilBERT Model Architecture Summary

3.8 Results and Interpretation

The model trainings, without the addition of the NLP dimensions as features, produced
correlation coefficients of around 0.73 for all LLMs. This would be equivalent to an R2 , or
goodness of fit, of 0.53, which would not make the models even fairly accurate in determining
the trustworthiness scores. The figures below show that all 3 LLMs with the additions
of the NLP dimensions had very good accuracy and goodness of fit, and the RoBERTa
model especially had the overall lowest MSE and the highest overall correlation. However,
some drawbacks of the RoBERTa model included the model accepting the least amount of
input tokens and requiring the longest training time. The addition of the NLP dimensions
significantly improved the accuracy of the models.

58
 Figure 3.18: Model Training Results

Figure 3.19: Model Validation and Testing Results

Figure 3.20: Model Correlation Coefficient Results

Despite these promising results, statistical significance tests needed to be performed between
combinations of the model scores and the true scores to validate if the RoBERTa model
would still be the best choice for the proposed data tool. Using the same subset of 1000
random observations that will be used for the unsupervised learning methods and AI model
evaluation, the real scores of those articles were compared with the predictions that the
BERT, RoBERTa, and DistilBERT models would make on that same data.

59
 Figure 3.21: Statistical Significance Test Results

Using T-tests and ANOVA to compare the means of these distributions, the mean of the real
score subset data was very close to the mean of the real scores of the overall dataset, which
means that these distributions would likely be normal. Every test showed that the means
of the sets of prediction scores and real scores were statistically significant as every p-value
was less than the designated value of α of 0.05. The ANOVA test showed that there was
a significant difference between the 3 model prediction score sets. The T-test comparison
that showed the least difference was between the real scores and the RoBERTa prediction
scores, and the T-test comparison that showed the greatest difference was between the BERT
prediction scores and the DistilBERT prediction scores.

Further validation of the results was done by using a subset of the WELFake Kaggle dataset
(https://www.kaggle.com/datasets/saurabhshahane/fake-news-classification) [75].
Using this subset, the scores that would be calculated using the NLP dimension algorithm
versus the scores that would be calculated using the RoBERTa model were compared with
the labels in the dataset. Note that the labels in the WELFake dataset were binary (0 and
1), so only classification accuracy could be used to compare these results. Although the ac-
curacies were low due to the NLP dimension vocabularies lacking complexity and real-world
examples found in the news today (these would likely improve the efficacy in the future),
these results showed that the RoBERTa model still performed better than the NLP dimen-
sion score calculations. These predictions were also statistically significant, meaning that

60
the RoBERTa model could still perform better compared to only using the NLP dimension
score calculations.

Figure 3.22: NLP Dimension Predictions vs. LLM Predictions Results

Therefore, the primary conclusion reached based on the model training and analysis is that
the RoBERTa model prediction scores follow the distribution and center of the real trust-
worthiness scores the most, making it the most efficient choice for the proposed data science
tool, while BERT would be the least effective choice as it had the lowest accuracy of the
three LLMs and the most disparity between its prediction scores and the real scores.

3.9 Model Integration in Proposed Data Tool

To integrate the chosen LLM into the proposed data science tool, the technique of extract-
ing the quantified NLP dimensions and concatenating them with the article texts was first
performed to create the inputs for the tool. This tool would require an input dataframe of
article texts and titles to generate the predictions effectively. The concatenated texts would
be encoded using the RoBERTa tokenizer and converted into input IDs and attention masks
to feed into the RoBERTa base model of the tool. Note that the exact model architecture
would be replicated as it was during training in order for the model prediction to func-
tion properly. The model would then generate the score predictions and in addition to the
predictions, the tool would output explainability metrics (i.e., presence of NLP dimension,
individual NLP dimension score, named entities list) to explain why the model likely gave
the provided score to the article text. Finally, the prediction scores would be stored in a new

61
column in the input dataframe and the dataframe with the predictions would be exported
to the user.

Figure 3.23: Diagram of Proposed Data Tool Structure

62
 Figure 3.24: Sample Proposed Data Tool Response

3.10 Summary

This chapter of the thesis explored multiple Hugging Face LLMs and their effectiveness in
determining trustworthiness scores for news articles regardless of length. Communication
elements of misinformation were converted into quantified NLP dimensions using indicative

63
vocabularies, and the inclusion of these dimensions into the LLM training significantly im-
proved the accuracy of the LLMs. These NLP dimensions also served as an important feature
to comprehensively label and assist in providing explainability for the score predictions.

After training all LLM options, RoBERTa had the best performance as its predictions fol-
lowed the distribution of the score labels the closest, despite it requiring the least amount of
input tokens or context. DistilBERT required the most input tokens and performed fairly
well, and BERT overall performed the worst as it had the least accuracy and the most dis-
parity between its predictions and the real scores. Therefore, RoBERTa was chosen to be
the base model for the proposed data tool.

Finally, a data science tool was proposed that would serve the purpose of efficiently gener-
ating score predictions through the base LLM by holistically evaluating the entire context
of the news articles using the NLP dimensions given an input dataframe of articles. This
tool would also create responses of explainability metrics that would assist with ensuring
trustworthiness in the tool by justifying why the given scores were generated for the news
articles. Likewise, score predictions can be stored and exported in dataframes provided by
the user.

Despite the evaluation of the fine-tuned supervised learning LLMs, unsupervised learning
models and AI models need to be compared to the performances of the LLMs to resolve
which of these 3 types of models is most recommended for future research and use for deter-
mining misinformation potential accurately.

64
Chapter 4

Unsupervised Learning Insights and Data

Visualizations

4.1 Overview

Unsupervised learning models will be the second type of model explored in this thesis. These
models do not require a labeled dataset nor training to make predictions, however they can
produce valuable insights about data including which clusters certain data points are a part
of. Given this application, these models can have the potential to produce novel revelations
about trustworthy and untrustworthy news articles.

The unsupervised learning methods that will be used for this portion of the thesis include:

• Method 1: Anomaly detection in tandem with LDA topic modeling.

• Method 2: Cosine similarity in tandem with t-SNE.

These 2 methods will be performed on a subset of the article dataset. Association rule mining
was also experimented with, however it was not be able to be used on the articles due to
GPU overload, and the last section of this chapter will briefly discuss related analysis that
was still conducted using this unsupervised learning method.

This chapter of the thesis will add to the conclusions reached from Chapter 3 and will attempt
to eliminate ambiguity between what makes a news article trustworthy or untrustworthy

65
through effective data visualizations and insights.

4.2 Data Collection

The following dataset will be used for the unsupervised learning analysis in this chapter:

• 1000 random observations from the training set used for the LLMs

Out of the 50,092 news articles in the training dataset, 1000 randomly extracted articles
will be used that follow the same score label distribution as the overall dataset. These
observations will be divided into 2 trustworthiness categories based on the score label range
that the articles fall into. These same 1000 observations will also be used in Chapter 5 of
this thesis for more effective and consistent comparisons.

4.3 Method 1

This approach involved anomaly detection and LDA modeling. The objective of this method
was to define an outlier subset of news articles and to distinguish the component differences
between the normal observations and anomaly observations between the articles that were
labeled "Likely Trustworthy" and "Likely Not Trustworthy". LDA models were also ini-
tialized to generate the top keywords and topics of the "Likely Trustworthy" and "Likely
Not Trustworthy" article anomalies. If the isolation forest model, which is responsible for
determining the anomalies that differ from the normal behavior of the data, would generate
anomalies that are trustworthy and not trustworthy articles, then the anomalies would be
most indicative of which articles are likely to be trustworthy and which articles are likely to
be not trustworthy. Although unsupervised learning methods do not require labeled data,
revealing patterns that distinguish trustworthy and untrustworthy news articles could pave
the way for unsupervised learning misinformation detection research that will not require
labeled data.

66
4.3.1 Data Preprocessing

For preprocessing related to anomaly detection, the 1000 observations were divided into 2
categories (depending on whether the labeled score was above 50 or below 50 as previously
explored in the training dataset in Chapter 3). This data would then be fed into the isolation
forest model.

The data preprocessing regarding the LDA modeling involved extracting the anomalies that
were labeled "Likely Trustworthy" and the anomalies that were labeled "Likely Not Trust-
worthy" to initialize the LDA models and generate wordcloud data visualizations for the
anomaly categories. All of the cleaned texts were tokenized and set to lowercase, and lists
of words and punctuation to exclude from analysis were defined that included insignificant
words and punctuation. It was important to ensure that as many prominent words as possible
were featured in the topics generated by the LDA model.

4.3.2 Anomaly Detection

The corpus of article texts was converted into a TF-IDF vector matrix, an acceptable input
for the isolation forest model. Then, the matrix had its features reduced using truncated
single value decomposition; the number of components, or features, to be analyzed was set to
2, therefore the anomalies would be analyzed on a 2-dimensional space. Then, the reduced
features were input into the isolation forest model with a contamination (proportion of data
that would be selected as outliers) of 0.1, so 10% of the data was chosen as anomalies to have
sufficient data from both trustworthiness categories for the anomaly detection analysis. In
the dataframe of article texts, anomalies were given a label of -1 to indicate that these were
anomalies; 62 articles were "Likely Not Trustworthy" anomalies and 38 articles were "Likely
Trustworthy" anomalies. Next, data visualizations were made comparing the anomaly data
to the normal data, which will be discussed in the Results and Interpretation subsection.

67
4.3.3 Latent Dirichlet Allocation Topic Modeling

For each trustworthiness category, the preprocessed article text anomalies were converted
into a dictionary and then into a corpus of text using bag-of-words. The LDA model would
be initialized using the dictionary, the corpus, and the following parameters:

• Number of topics: 3

• Number of passes: 10

• α: 0.7

• η: 0.7

After the LDA model was trained to perform the topic modeling, 3 keywords were generated
for each of the 3 topics specified by the model.

4.3.4 Results and Interpretation

The results of the anomaly detection and LDA modeling, especially through the use of data
visualization, revealed some enlightening insights. Figure 4.1 shows the anomaly data using
purple dots and the normal data using yellow dots, displayed in a 2-dimensional space that
represents the 2 components, or reduced features of the data.

68
 Figure 4.1: 2-Dimensional Visualization of Anomalies

(a) Component 1 (b) Component 2

Figure 4.2: Boxplot Visualizations of Anomalies

Based on the visualizations of the components comparing the normal and anomaly data in
Figures 4.1 and 4.2, these show that the normal data appear to be concentrated within the
range of 0.15 and 0.6 for the value of Component 1, and concentrated within the range of -0.1

69
and 0.1 for the value of Component 2. In contrast, the anomaly data tend to have a larger
range, with 0.1 to 0.75 for Component 1 and -0.1 to 0.9 for Component 2. The centers of the
boxplot distributions for the normal and anomaly data appear to be the same at about 0.35
for Component 1, but the centers of the distributions for Component 2 are more different,
with the normal data centered around 0.0 and the anomaly data centered around 0.1.

(a) Component 1 (b) Component 2

Figure 4.3: Distribution Visualizations of Anomalies

Density distribution visualizations likewise provided crucial content for analysis. Figure
4.3 shows the normal data distributions being unimodal for both components while the
anomaly distributions for both components are bimodal. Given this insight, additional
analysis was done to see which trustworthiness category was more prevalent in the anomaly
data distributions given specified thresholds of the component values. It was found that
likely trustworthy articles were twice as prevalent compared to likely untrustworthy articles
with a Component 1 value greater than 0.5, and that likely not trustworthy articles were
three times as prevalent with a Component 1 value less than 0.3. Likely not trustworthy
articles were almost 7 times more prevalent with a Component 2 value greater than 0.5, and
likely trustworthy articles were almost 2 times more prevalent with a Component 2 value
less than 0.05. Therefore, trustworthy anomalies tended to have a higher Component 1
value and a lower Component 2 value, while untrustworthy anomalies tended to have a lower

70
Component 1 value and a higher Component 2 value.

When using a contamination of 0.05, the results using the initial contamination of 0.1 changed
some; regardless of component values, "Not Likely Trustworthy" anomalies were more present
within each of the lower and higher value ranges of the distributions. However, "Not Likely
Trustworthy" articles were 4 times more prevalent than "Likely Trustworthy" articles in
the anomalies determined by the isolation forest model (39 to 11). Thus, while having a
lower contamination can lead to one trustworthiness category likely being isolated, a higher
contamination can provide further differentiations between the 2 trustworthiness categories
through their component values.

Figure 4.4: Topics and Keywords in Likely Trustworthy Anomaly Articles

Figure 4.5: Topics and Keywords in Likely Not Trustworthy Anomaly Articles

The topics and keywords generated by the LDA models on the likely trustworthy and likely
untrustworthy category data showed that articles discussing topics such as former U.S. pres-
ident Donald Trump, the Wagner Group and Prigozhin in Russia, and the U.S. Senate and
court system are likely trustworthy, and articles discussing topics such as NIST websites and
links and people and identity are likely not trustworthy.

Wordcloud data visualizations were also created to illustrate the most significant words
present in the anomaly data for each trustworthiness category. The article texts were pre-

71
processed the same way as they were for the LDA modeling, however a counter was also
used for counting the number of times a given word appeared in the corpus of text. The
wordcloud displayed all significant words that were present at least 10 times in the corpus.

Figure 4.6: Likely Trustworthy Article Words

Figure 4.7: Likely Not Trustworthy Article Words

The wordcloud visualizations contain some of the same words that were shown in the gener-

72
ated topic keywords from the likely trustworthy and likely untrustworthy articles. In addi-
tion, words that indicate articles discussing topics including politics, public officials, former
presidential candidate Hillary Clinton, government, the war in Syria, and web vulnerabilities
and security are articles that are likely to be trustworthy. Articles that include words such
as information, content, bash, endorse, think, and page indicate that these articles are likely
to be not trustworthy. There are some words that are present in both wordclouds, which
still leaves some ambiguity between trustworthy and untrustworthy news article keywords.

Therefore, based on this unsupervised learning approach, this analysis has shown that certain
keywords as generated from the LDA modeling and wordcloud visualizations can provide in-
dications for whether a news article is likely to be trustworthy or untrustworthy based on its
topic. However, there can be some overlap between words occurring in both trustworthiness
categories. Furthermore, this method has shown that if using dimension reduction through
anomaly detection on news articles (set to 2 dimensions only), an article with a higher value
for Component 1 was more likely to be trustworthy, and an article with a higher value for
Component 2 was more likely to be untrustworthy.

4.4 Method 2

This approach involved cosine similarity and t-SNE. The objectives of this method were to
determine if there were distinctions between the cosine similarities of the likely trustworthy
news articles and the likely untrustworthy news articles and to create visualizations to see
if certain clusters of likely trustworthy news articles and likely untrustworthy news arti-
cles tended to separate from each other. Component distributions for each trustworthiness
category were also analyzed to see if there were any notable differences. If any significant
distinctions would be revealed between the trustworthiness categories, this method would
make an effective indicator to conclude which news articles are trustworthy or not in future
misinformation detection research.

73
4.4.1 Data Preprocessing

The same 1000 observations were used and divided into the same trustworthiness categories
as in Method 1. Additionally, the categorized news articles were combined and preprocessed
using NLTK tokenization, setting all tokens to lowercase, and removing stopwords and punc-
tuation. This cleaned data would then be fed into the Word2Vec model that would then be
used to calculate the cosine similarities between the news articles in vector form.

4.4.2 Cosine Similarity

The news articles texts were converted into word embeddings, or vectors, using a Word2Vec
model in the Gensim package, and each vector was set to a size of 200. Then the model
vocabulary was constructed and the model trained. Document vectors for the articles were
generated by iterating through the processed articles to create a single vector for each article.
The cosine similarities comparing the document vectors were calculated and put into vector
form.

4.4.3 t-Distributed Stochastic Neighbor Embedding

Taking the document vectors, these were put into a NumPy array and any remaining missing
document indices were dropped. This array was then transformed into embedded vectors
using t-SNE with the following hyperparameters:

• n_components: 2

• perplexity: 20

• random_state: 42

Therefore, the t-SNE analysis will be done on a 2-dimensional scale.

74
4.4.4 Results and Interpretation

Using the cosine similarity vectors generated, the averages of the vectors were calculated
and a violin plot was made comparing the average cosine similarities of each trustworthiness
category. Figure 4.8 shows that the distributions of the average cosine similarity for each
trustworthiness category were approximately the same, with the exception of differences
between the amounts of outliers.

Figure 4.8: Violin Plot Comparing Average Cosine Similarities

In addition, a statistical t-test was conducted to see whether the average cosine similarities of
the trustworthiness categories were significantly different. The t-test calculated the following:

• T-Statistic: 0.08917

• P-Value: 0.92895

Thus, the average cosine similarities between the two trustworthiness categories were not
significantly different given a significance value of α at 0.05.

75
Next, data visualizations depicting the 2 dimensions of the cosine similarities using t-SNE
were made. Figure 4.9 shows some overlap between the two trustworthiness categories,
however some distinct clusters included likely not trustworthy articles with dimension values
around X = -30 and Y = 40, and likely trustworthy articles with dimension values around
X = -50 and Y = 0. As seen in Figure 4.10, the t-SNE dimension distributions appeared to
be approximately the same, however the likely untrustworthy news articles tended to have a
more bimodal distribution for both dimensions. Finally, Figure 4.11 shows that the parallel
coordinates between the 2 dimensions of both trustworthiness categories overlapped as well.

Figure 4.9: t-SNE Plot Comparing Trustworthiness Categories

76
 (a) Dimension 1 (b) Dimension 2

Figure 4.10: Distribution Visualizations of t-SNE Dimensions

Figure 4.11: Parallel Coordinates Plot

Using this unsupervised learning method, analysis showed that while there can be some
differences between the cosine similarities and t-SNE dimension values of trustworthy and

77
untrustworthy news articles, there was a lot of overlap between the dimension distributions
of the trustworthiness categories such that these categories could not be significantly distin-
guishable. Therefore, this method would not be effective for future misinformation detection
research and would not be reliable to help combat misinformation on a big-data scale.

4.5 Summary

The results have shown that using Method 2 there were no significant differences between
the cosine similarities, nor differences between the dimension distributions of the trustworthy
and untrustworthy news articles, thereby it is an inefficient method for misinformation detec-
tion. In contrast, Method 1 was more effective at differentiating between likely trustworthy
and likely not trustworthy news articles, however there was still some overlap between the
distributions of likely trustworthy and likely untrustworthy articles. Method 1 still revealed
some enlightening insights including certain reduced dimension values and topic keywords
that can help eliminate ambiguity in what makes a news article trustworthy or not, and it can
be encouraged to pursue this method for future research as unsupervised learning methods
for misinformation detection have generally been understudied. However, it is questionable
whether unsupervised learning methods are effective enough to be suitable for open-source
use given the very strong results that the LLMs in Chapter 3 provided.

So far in this thesis, in the context of misinformation detection using news articles, un-
supervised learning models have not been consistent in producing effective results while
fine-tuned supervised learning LLMs have been consistent. Futhermore, both unsupervised
learning methods still tend to have some overlap between the trustworthiness category distri-
butions, which makes unsupervised learning not an absolute candidate for the most efficient
data model for misinformation detection. Although the evaluation of AI models is yet to
be discussed, the fine-tuned supervised learning LLMs should still be regarded as the most
effective models at this point in the thesis.

78
4.6 Note on Association Rule Mining and Related Work

Consistency between the data evaluated but using different models on that data is the
most suitable approach for determining which type of model is most recommended for
future misinformation detection research and use by target stakeholders, therefore this
explains why the same news article data is being utilized for all evaluations. Associa-
tion rule mining was unfortunately not implemented in this thesis using the article data
due to complications involving GPU availability, and this was likely due to the length
of the news articles generating too many frequent itemsets and association rules. How-
ever, there was a separate analysis that was able to be conducted using a verified true
and false statement dataset called the Politifact Fact Check Dataset from Kaggle (https:
//www.kaggle.com/datasets/rmisra/politifact-fact-check-dataset). In this experi-
ment, association rule mining and Methods 1 and 2 were practiced on this dataset; the results
yielded more distinctions between the verdict categories of the statements (true, mostly-true,
half-true, mostly-false, false, and pants-fire) for all unsupervised learning methods. This work
was especially insightful because the statements in the dataset can be classified as objective
or unintentional misinformation, which is very hard to distinguish from true facts, therefore
this work provided some practical results that could contribute to automated fact-checking
research and could have the potential to be publishable work in the future. Data visualiza-
tions that resulted from this analysis can be found in Chapter 10. Although it is important
to acknowledge this work and its contribution to the domain of misinformation, given the
broader context of this thesis it would not be recommended to digress to discuss a modeling
approach on a dataset inconsistent with the data being used for the 3 components of the
thesis.

79
Chapter 5

Evaluation of Artificial Intelligence Model

Usage for Misinformation Detection

5.1 Overview

The final type of model that will be explored for this thesis will be generative AI models.
These models are designed to provide answers to a variety of questions prompted by the
user, and these models can have the capacity to return a trustworthiness score for a given
news article text. As these models are also accessible in pay-as-you-go APIs, the user can
also have the opportunity to store the API responses in a dataframe or database.

Given this very advanced functionality, it could be (harmfully) assumed that AI models
like ChatGPT can be used for automated misinformation detection in its API form. While
automated methods are greatly needed for misinformation detection, the recent arrival of
such generative AI models into this domain should not be taken for granted as reliable
solutions. There is potential that these models may have bias or inaccuracy in their decision-
making, and if these models would be used for misinformation detection given these flaws,
these models could send a distorted perception to users, which would make these users
reliable on a tool that provides inherently wrong or misguided judgments of misinformation
potential. Therefore, this is why these models need to be evaluated in order to ensure which
type of model is most suitable for accurately determining misinformation and trustworthiness

80
in news articles.

5.2 Data Collection

In Python, OpenAI and Google have packages that are required to use their AI model APIs,
and these packages have their respective chat functions to access these APIs. Unique API
keys provided by their respective portals or administration were also required to access the
APIs. The chat functions can have some parameters specified by the user, including the
type of model (ChatGPT-3.5 and 4 each have multiple versions of the models that can vary
in expense), the temperature, or consistency, of the response, and the desired prompt. As
of January 2024, OpenAI’s pricing for its API models has changed, however the models
that were used at the time of analysis (August-September 2023) with their pricing will be
referenced. Figures 5.1 and 5.2 show the example setups of the ChatGPT and PaLM API
prompts to collect the trustworthiness scores.

Figure 5.1: ChatGPT API Prompt Code

Figure 5.2: PaLM API Prompt Code

After setting up the prompts, each news article text concatenated with the prompt was
passed through the API, which returned a response for each news article text explaining the

81
trustworthiness score given to the text and reasons why. These responses were stored in the
dataset and later the scores contained in the responses were parsed and stored in another
column in the dataset. Sometimes the API could return a response but not a definitive
score for the text; after testing to see if the responses would change for these texts, it was
concluded that the model usually could not determine scores for these specific texts and
therefore these texts had to automatically be given a score of 0.

5.3 Application Programming Interface Setup

5.3.1 ChatGPT-3.5

ChatGPT-3.5 utilized the entire dataset used in Chapter 3 as its Turbo version of the model
only cost $0.0015 per 1000 input tokens, which was relatively cheap. Note that for Chat-
GPT’s API output tokens are also charged with use, however there were definitively not as
many output tokens compared to input tokens while using the API.

For collecting the ChatGPT-3.5 trustworthiness score data, the following parameters were
set:

• model: gpt-3.5-turbo

• temperature: 0.1

• role: user, content

• message: Rate the trustworthiness of this article text on a scale from 0 to 100, with 0
being not trustworthy and 100 being most trustworthy: article text

82
 Figure 5.3: Sample Response from ChatGPT-3.5 API

5.3.2 ChatGPT-4

ChatGPT-4 did not have a Turbo model in existence at the time of analysis, so its API model
cost $0.03 per 1000 input tokens, which was 20 times more expensive than ChatGPT-3.5.
Due to this very high expense, a random sample of 1000 news articles was used for analysis.
These were also the same 1000 news articles used for the unsupervised learning methods and
LLM comparisons.

For collecting the ChatGPT-4 trustworthiness score data, the following parameters were set:

• model: gpt-4

• temperature: 0.1

• role: user, content

• message: Rate the trustworthiness of this article text on a scale from 0 to 100, with 0
being not trustworthy and 100 being most trustworthy: article text

83
 Figure 5.4: Sample Response from ChatGPT-4 API

5.3.3 PaLM

The PaLM API was free to use, however the API’s response time for passing API calls was
much slower than ChatGPT-3.5 and 4, therefore the same 1000 observations were used for
collecting the PaLM scores as well.

For collecting the ChatGPT-4 trustworthiness score data, the following parameters were set:

• model: chat-bison-001

• temperature: 0.2

• candidate_count: 1

• top_k: 40

• top_p: 0.95

84
 • messages: Rate the trustworthiness of this article text (numerically) on a scale from 0
to 100: article text

Figure 5.5: Sample Response from PaLM API

5.4 Comparative Analysis and Interpretation

5.4.1 Reasons for Scoring

This subsection will discuss the results comparing the generative AI models based on ex-
planations for scoring. First, the reasons given by the AI models for providing their trust-
worthiness scores need to be discussed. Reasons why the ChatGPT-3.5 API provided a low
score for a news article text included:

• Not enough context or information that could be fact-checked (these were the texts
that could not be given a score by the API).

• A lack of verifiable sources to support certain claims made in the article.

• Strong biased language.

• Irrelevancy of information.

• Lack of specific details.

• Promoting the author’s personal opinions instead of objective information.

Reasons why the ChatGPT-3.5 API provided a high score for a news article text included:

• Specificity of information.

85
 • Direct quotes from persons of interest in the article.

• Inclusion of credible sources and expertise.

• Citation of sources.

• Factual information.

• Inclusion of multiple perspectives.

Overall, ChatGPT-3.5 did an efficient job of providing trustworthiness scores with in-depth
explanations of factors that led to the determination of these scores. There were, however,
some aspects that should be noted about its performance, including its inability to provide
trustworthiness scores for some texts if they do not include enough context or information to
fact-check. Another limitation to note is that it acknowledges it cannot serve as an absolute
source for verifying news article texts as it is still important for the user to cross-reference
information in the article with other sources in order to confirm the article’s validity.

Next, to discuss ChatGPT-4’s effectiveness at providing trustworthiness scores with expla-

nations, reasons why it provided a low score for a news article text included:

• Heavily biased, sarcastic, or inflammatory language.

• Lack of sources to support claims made in the article.

• Generalized, assumptive, or speculative claims.

• Informal or disjointed writing style.

Reasons why the ChatGPT-4 API provided a high score for a news article text included:

• Detailed information, data, and statistics.

• Reputable sources.

• Well-written writing style.

86
 • Inclusion of additional context.

• Comprehensive overview and explanation of the topic of interest.

• Real interviews with credible persons of interest.

• Direct quotes from experts or persons of interest.

However, explanations with the dimensions of trustworthiness scoring that ChatGPT-4 pro-
vided were a minority of responses from the API. Most of the responses consisted of only the
score with no explanation. Although the API calls appeared to be returned efficiently with
some extensive responses, ChatGPT-4’s ability to not provide explanations for the scores for
a majority of the time raises questions about its transparency for misinformation detection.

Finally, reasons why the PaLM API provided a low score for a news article text included:

• The article text originated from an unreputable news outlet such as "The Onion" or
"The Daily Caller" (however this was not true about the actual origin of the article).

• Unidentified author.

• Unknown or not widely known publication.

• Specific false claims about vaccines, "the Earth is flat", etc. were made in the article
(this was also not true about the actual content of the article).

• Sensationalist tone.

• Lack of evidence to support claims.

Reasons why the PaLM API provided a high score for a news article text included:

• Well-written text.

• Clear and concise overview of the topic.

87
 • Evidence to support claims.

• Free of grammatical errors.

• Inclusion of expertise.

Despite the PaLM API appearing to provide scores and detailed explanations for the news
articles, there were several elements in the responses that were extraordinarily flawed and
inaccurate. First, the API tended to provide repetitive responses that mentioned that trust-
worthiness depends on a number of factors including author credentials, publication reputa-
tion, and article content. Some additional dimensions were included in the score explanations
but the AI model tended to evaluate the article texts based on those 3 factors, which may
not necessarily be practical for comprehensively determining a trustworthiness score solely
based on an article text. It also acknowledges that news article content should still be veri-
fied through cross-referencing with additional information.

Next, this AI model also had a strong tendency to misinterpret the article origin and/or
content, which was very alarming to discover. Figure 5.6 shows an article text in which
its content covered the topic of an information leak related to the Department of Defense,
but the PaLM API response says that the article covered the topic of the benefits of eating
chocolate and was written by a certified diabetes educator. Another example, Figure 5.7,
shows a response that says the article originated from the satirical news outlet "The Onion";
the actual article origin was The Business Journals. In contrast, the ChatGPT API was able
to correctly and consistently reference information given in the prompt in its responses.

Figure 5.6: Example of Misinterpreted Article Content

88
 Figure 5.7: Example of Misinterpreted Article Origin

To interpret these results, out of these 3 AI models ChatGPT-3.5 had the most superior
capability to provide extensive, multi-faceted explanations for trustworthiness scores. Al-
though ChatGPT-4 was sometimes able to generate explanations for trustworthiness scores,
its ability was definitively not as consistent nor transparent compared to ChatGPT-3.5’s.
Furthermore, PaLM had very fallacious explanations for its trustworthiness scores as it
tended to severely misinterpret crucial aspects of the article texts. What could explain these
illogical results with PaLM is due to the API being free to use (thereby meaning potentially
less quality in responses) or due to the API having limited access at the time of data collec-
tion (meaning that the API was in its "testing" phase with a restricted public user clientele).
Lastly, some of the reasons that the models provided in their explanations aligned with the
communication dimensions of misinformation that were used to create the trustworthiness
score labels for the article texts, meaning that there is some legitimacy in their explainability.

5.4.2 Accuracy in Scoring

This subsection will discuss the results comparing the generative AI models based on accu-
racy in scoring. Evaluating the accuracies of the AI models was conducted by comparing
the numerical differences between the model scores and the labeled scores as well as us-
ing binary classification (assigning 0 to all scores below 50 and assigning 1 to all scores
above 50). Although using supervised learning was a regression-based task and not really a
classification-based task, it was still important to determine whether the AI models tended
to be consistent in predicting whether news articles were likely to be trustworthy or likely
to be not trustworthy.

The differences between the scores were calculated such that each labeled score was sub-

89
tracted from each corresponding AI model score, therefore if the difference was positive that
meant the AI model tended to score higher. As shown in each of the distributions in the
histogram below, the distribution tended to skew left, which meant that the average score
difference for each of the AI models was positive. The median ChatGPT-3.5 score difference
was 16, the median ChatGPT-4 score difference was 31, and the median PaLM score dif-
ference was 17. Thus, it is shown that the AI models tended to inflate the trustworthiness
scores of news article texts to make it appear that these trustworthiness texts are likely to be
more trustworthy when these texts may actually not be trustworthy, which is a potentially
concerning finding.

Figure 5.8: Normalized AI Model Score Differences Histogram

To further evaluate the numerical differences between the scores, a scatter plot comparing
all 3 AI model predictions was made to see if there were any correlations between the labeled
scores and the AI model scores. While the scatter plot is eccentric in display and did not
reveal any significant correlations between the scores, this plot still showed that the AI
models tended to allocate higher trustworthiness scores, as shown by more rows of data
points near the top of the Y-axis.

90
 Figure 5.9: Labeled Scores vs. AI Model Scores

The R2 , or correlation, for each set of AI model predictions compared to the labeled scores
was computed. These correlations were then compared with the previous correlation results
using the LLMs; because there were many higher scores from the AI models, and because
the scores were not continuous, the R2 s for the AI models were very low. The scatter plot
in Figure 5.9 also supports these results. ChatGPT-3.5 still had the highest correlation
coefficients and R2 while PaLM had the lowest; in general, between comparing the LLMs
and the AI models, RoBERTa would still be the most efficient choice.

Figure 5.10: Correlations for AI Models Compared with LLM Correlations

91
Statistical significance tests were likewise conducted to ensure the differences between the
3 distributions of the AI model scores. Since these distributions were not normal, non-
parametric tests including the Kruskal-Wallis Test and the Mood’s Median Test were con-
ducted to compare the medians. Using a value of α at 0.05, the tests showed the following
results, which showed that these score distributions are indeed significantly different:

• Kruskal-Wallis Test Statistic: 1198.9978735287934

• P-value: 4.374413733457815e-261

• Mood’s Median Test Statistic: 749.63889942343

• P-value: 1.6518905936712285e-163

Although the data in the 3 AI model score distributions are not normal, it is important
to note how close the centers of these distributions are to the center of the actual model
score distribution. The mean value of the labeled scores, ChatGPT-3.5 scores, ChatGPT-4
scores, and PaLM scores are respectively as follows: 52.62, 62.97, 81.45, 67.15. The median
value of the labeled scores, ChatGPT-3.5 scores, ChatGPT-4 scores, and PaLM scores are
respectively as follows: 53, 70, 85, 70. Comparing these measures of center, it can be
concluded that ChatGPT-3.5 overall tended to have a closer center of distribution to the
center of the actual score distribution.

Next, regarding the classification accuracy of the AI models, the transformed binary labeled
scores and AI model scores were used for this analysis. Data visualizations including Area
Under Receiving Operating Characteristic (AUROC) curves and confusion matrices that are
commonly used for testing classification accuracy were created as well.

The AUROC was calculated for each AI model, and the ROC curves for the models are
shown in Figure 5.11. ChatGPT-3.5 had the highest AUROC of 0.56, however all 3 AI
models had AUROCs of around 0.5-0.6, which generally means that these AI models are no

92
better than random chance at distinguishing which news articles are likely to be trustworthy
and which articles are likely to not be trustworthy.

Figure 5.11: AUROC Curves of AI Model Scores

Confusion matrices were also made for comparing how the AI models performed for cor-
rectly identifying likely trustworthy and likely untrustworthy articles. While the AI models
were more likely to correctly predict likely trustworthy articles, they were also more likely
to incorrectly predict likely trustworthy articles, thus leading to a significantly large false
positive rate. All AI models’ false negative rates were relatively low but still present. These

93
results support the previous conclusions made from analyzing the score differences (i.e., a
false positive is an indicator that the AI model adds more trustworthiness value to a news
article when the actual trustworthiness value is lower). The high false positive rate should
also serve as a warning that AI models would tend to inform a user that an article text
is likely to be trustworthy when it actually may not be, thereby it is concerning to use AI
models for open-source misinformation detection.

Figure 5.12: Confusion Matrix of ChatGPT-3.5 Scores

Figure 5.13: Confusion Matrix of ChatGPT-4 Scores

94
 Figure 5.14: Confusion Matrix of PaLM Scores

Finally, confusion matrix metrics were calculated for each AI model to determine which
model had the best effort in correctly identifying the likely trustworthiness of the news
articles. As seen in Figure 5.15, all 3 AI models had an accuracy of just over 0.6, however
ChatGPT-4 had the highest accuracy, precision, recall, and F1-score. All 3 AI models’ recalls
were significantly high due to their low false negative rates. Although ChatGPT-4 had the
best performance in accurately predicting trustworthiness classification, it is important to
note that the ChatGPT-4 model was used on a much smaller subset of the dataset; if the
model was used on the entire dataset, there is a potential chance that the accuracy could be
affected.

Figure 5.15: AI Model Classification Metrics

These were compared with classification accuracy metrics using the RoBERTa model in

95
Figure 5.16. Although the RoBERTa classification accuracy was roughly the same as the
other AI models, the lower accuracy compared to the strong testing R2 could be due to the
high number of labeled scores in the middle score range (45-55) in which the threshold of 50
set for the classification task would have starkly divided these articles into separate categories
despite these articles having minimal differences between their actual trustworthiness scores.

Figure 5.16: RoBERTa Classification Metrics

Therefore, comparing the AI models versus the well-performing RoBERTa model, despite
the lower classification accuracy of the RoBERTa model, it still had the best performing
testing R2 . Classification accuracy results using RoBERTa could be improved in the future
by making the distribution of the labeled scores more bimodal through NLP dimension
adjustments. However, given the results of the AI models from this chapter, which imply
insufficient explainability and accuracy to make them reliable for open-source use, these
would not be recommended for misinformation detection tasks. Results using the AI models
could likewise be improved through prompt engineering; although more scores could be
collected from the APIs using more directed prompts, it is difficult to determine whether
this would improve the accuracy of the results.

5.5 Summary

The ChatGPT-3.5 model did the best job at explainability for predicting trustworthiness
scores and also had the highest AUROC, while the ChatGPT-4 model had the highest
confusion matrix accuracy. However, given the additional context that ChatGPT-4 did not
consistently provide detailed explanations for the trustworthiness scores and that it also
had the highest median score difference between its scores and the corresponding labeled

96
scores, it would not be a recommended AI model to use for future misinformation detection.
Likewise, it would strongly not be recommended for PaLM to be used for misinformation
detection as it produced very inaccurate responses and had the least effective accuracy in
scoring.

Although ChatGPT-3.5 would be the best candidate out of the 3 AI models tested for
automated misinformation detection, it still tended to inflate the values of the trustworthiness
scores (as did the other AI models), which compared to the LLMs and unsupervised learning
methods, this would be concerning to use for crucial tasks like misinformation detection if the
actual trustworthiness values of news articles are lower than the AI model says. Likewise, a
64.11% classification accuracy for ChatGPT-3.5 is not sufficient evidence to conclude that it
would be reliable for misinformation detection by public users; if the classification accuracy
was at least 80%, then it is possible that this model could be recommended for misinformation
detection in the future.

It should also be noted that more studies will be needed to further evaluate all 3 AI models in
their misinformation detection abilities, therefore these models can be used with caution for
research purposes, but based on the results in this chapter they should not be recommended
as automated open-source alternatives.

97
Chapter 6

Discussion

6.1 Implication of Findings to Research Problems

Evaluating all 3 types of misinformation detection models for this thesis has yielded several
practical insights that will contribute greatly to the data science field as well as to the
misinformation communication domain. This section will discuss the significance of the
findings of each component of the thesis to general knowledge and how they addressed
the research questions presented in Chapter 1. Limitations of the thesis work will also be
presented with potential solutions to these problems in later sections of this chapter.

6.1.1 Data Tool Using Large Language Model

To begin, the primary research question pertaining to Component 1 research for this thesis
was:

Can a data science tool that rapidly and accurately assesses the misinformation
likelihood of news articles, while ensuring transparency through incorporating
key NLP dimensions, be developed to assist stakeholders who are potentially
impacted by, or may have a role in, combating misinformation?

The findings from Component 1 showed that supervised learning LLMs were very efficient at
producing accurate results using a regression-based trustworthiness scoring framework with

98
included quantified NLP dimensions of misinformation. The 3 Hugging-Face LLMs tested
(BERT, RoBERTa, and DistilBERT) consistently showed strong results; although it required
the least amount of input tokens, RoBERTa performed the best with a correlation coefficient
of around 0.995-0.996. To address the research question, the saved weights for the model
could be saved and used, while reproducing the architecture of the model, to create auto-
mated and holistic trustworthiness score predictions while at the same time incorporating
quantified NLP dimensions utilizing the news article content that could not be able to be
included in the input tokens that would be put into the LLM. This LLM would then serve as
the base for the proposed data science tool by incorporating the NLP quantified dimensions,
the score predictions, and the explainability of the NLP dimensions to justify why the tool
provided those scores.

The finding that a data science tool can indeed be created for the purposes and type of
information data mentioned in the research question can provide an opportunity to con-
tribute to the diversity of available tools used for misinformation and fake news detection.
As explainable AI and machine learning is likewise a pressing topic in current data science
research, this tool will also contribute to transparency in misinformation detection models.
The NLP dimensions served an extremely significant and versatile role in labeling data,
quantifying communication dimensions of misinformation, improving model accuracy, and
providing explainability in the proposed data tool, therefore the addition of these features
helped immensely in providing quality to the model predictions. The addition of the NLP
dimensions did likewise contribute an effective way to bypass the limitation of the LLMs
requiring a limited number of input tokens while still making strongly accurate predictions;
the chosen RoBERTa model for the data science tool required 160 input tokens (even less
than the maximum 512) as going over that would cause GPU overload in Google Colab. The
proposed tool simultaneously incorporating the extracted quantified NLP dimensions into
the score predictions created an effective data science approach to comprehensively evaluate
the entire content of news articles despite the LLM input token limitations.

99
The second research question that Component 1 should also address was:

Which dimensions of misinformation will be considered when creating the data

science tool for scoring the trustworthiness of a news article?

13 communication dimensions of misinformation were proposed based on the variety exam-

ined in the literature review, and while some dimensions may not always be present in news
articles, all should be regarded as individual holistic measures of trustworthiness to deter-
mine whether a news article is trustworthy or not. Each of these dimensions was assigned
a weight such that depending on the presence or absence of a dimension, the highest pos-
sible cumulative trustworthiness score would be 100, while the lowest possible cumulative
trustworthiness score would be 0. All of these dimensions were quantified using specified
indicative vocabularies and used as additional NLP features in the LLM, which as previ-
ously mentioned the addition of these dimensions to the LLM did produce strongly accurate
results.

The contribution of these proposed dimensions can provide a different perspective to previous
interpretations of which communication dimensions of misinformation should be considered
most important. This is especially important for determining which textual dimensions of
misinformation should be considered if the model should only evaluate based on text input,
rather than also having to incorporate other factors including author, publication credibility,
etc. in which that data may be more difficult to collect and compile. Communication dimen-
sions converted into NLP features can also improve the accuracy of future misinformation
detection models used in data science research.

6.1.2 Unsupervised Learning Approaches

The third research question, regarding Component 2, was:

Can unsupervised learning methods performed on trustworthy and untrustwor-

100
thy news articles provide any valuable insights for future research on misinfor-
mation detection?

The research findings included that the unsupervised learning approach using the techniques
of anomaly detection and LDA topic modeling did show some promising insights, including
that when anomalies were extracted from the dataset using dimensionality reduction, which
news articles were likely to be trustworthy and which articles were likely to be untrustworthy
could be determined based on the reduced component values in the anomalies. This approach
also showed that generated topic keywords could likewise be indicative of articles that were
likely to be trustworthy or articles that were likely to be untrustworthy. In contrast, how-
ever, the approach using the techniques of cosine similarity and t-SNE could not produce
any practical insights as the distributions of the component values for the trustworthy and
untrustworthy articles overlapped greatly, thereby not creating much distinction between the
trustworthiness categories. In addition, the technique of association rule mining was unable
to be experimented on the news article data due to GPU overload.

Therefore, only one method that was discussed in this thesis was able to provide decently
practical insights that can be very useful for misinformation detection research. The findings
from Method 1 can contribute to general knowledge such that if research is done using news
article datasets for fake news detection, converting these news articles to reduced dimension
values can determine which news articles are likely trustworthy and which ones are likely
not based on their dimension, or component, values. Likewise, LDA topic modeling has pro-
vided some practical topic keywords and data visualizations that can help assist with future
research by providing some preliminary topics that may indicate a likelihood of whether a
news article is trustworthy or not. Therefore, this technique can contribute to the array of
potential options for less-studied unsupervised learning approaches used for misinformation
detection.

101
6.1.3 Generative Artificial Intelligence Model Evaluation

The final research question posed in Chapter 1 of the thesis was:

Do generative AI models have the potential to detect misinformation accurately?

The findings from Component 3 showed that OpenAI’s ChatGPT-3.5 model did the best job
of providing explainability for trustworthiness scores, had the highest AUROC, and had the
closest mean trustworthiness score to the mean of the actual labeled scores, but did not have
the best classification accuracy. In contrast, Google’s PaLM model did not provide accurate
explainability for the trustworthiness scores and had the lowest AUROC and classification
accuracy. The ChatGPT-4 model had the highest average score difference from the actual
labeled scores and also did not provide much explainability for its scores in its responses.
However, all 3 AI models (especially ChatGPT-4) inflated the values of the trustworthiness
scores, which raises the concern that the AI models are either interpreting the news article
texts inaccurately and scoring them higher than they should be scored, or even more danger-
ous, they are providing false information in their responses as to make the trustworthiness
of the news articles higher than they actually are based on cherry-picked factors present in
the article.

Addressing the research question, based on the findings from Component 3 it is inappropri-
ate to conclude that generative AI models have the potential to serve as accurate automated
misinformation detection models. Although ChatGPT-3.5 had the best results out of the 3
tested, it still tended to overestimate its trustworthiness scores and its average trustworthi-
ness score was still significantly different from the labeled scores’ mean score. As the average
classification accuracy among the 3 models was around 64%, this is likewise not a sufficiently
strong accuracy to direct a conclusion that AI models are reliable for misinformation detec-
tion tasks; if the average accuracy was much higher (at least 80-90%), that conclusion could
change. Unless future studies show otherwise, these results produced from Component 3

102
are currently consistent with results produced from literature review sources analyzing AI
models’ efficacy in fake news detection [7], [28].

The findings from Component 3 should serve as a preliminary study of open-source AI mod-
els’ ability to detect misinformation, which these models have proved to be untrustworthy
in their ability to be accurate and reasonably explainable. While more data science field
research will likely be needed to further evaluate these models for misinformation detection,
these models so far do not hold promising results and should be avoided for automated
misinformation detection (unless for field research purposes only).

6.1.4 The Most Efficient Model

Given the contributions of the 3 components of this thesis, there is a final supplementary
question that should be addressed:

To conclude this work, which type of model is the most consistently efficient at
misinformation detection: fine-tuned supervised learning LLMs, unsupervised
learning models, or AI models?

To give a concise summary, all 3 LLMs gave very accurate results, only 1 unsupervised
learning approach gave sufficiently promising results, and none of the tested AI models gave
convincingly accurate results in misinformation detection. Given this comparison, it should
be concluded that training a supervised learning model, especially a fine-tuned LLM with
the ability to capture textual context, still stands as the most effective type of approach to
accurately determine misinformation and fake news. Out of the 3 LLMs tested, RoBERTa
produced the best results, which could likely be attributed to its more robust training nature,
therefore if an individual model had to be chosen to pursue for future misinformation detec-
tion tasks, this model would be most recommended. Likewise, since the LLM has proved to
be the superior model type, it can still function as a very effective and practical base model
for the proposed data science tool in this thesis.

103
6.2 Limitations

Despite the potential for progress in contributing to the misinformation detection domain,
there are some limitations of this thesis work that need to be noted. The most significant
limitation of this work, which can have the most impact on how valid this work may be
to use in the future, is that the vocabularies used for encoding the NLP dimensions were
vague in some qualities. None of the dimension vocabularies had indicative phrases spanning
more than one word included. Since tokenization only involves single words or puncutation
marks, while it is easy to count the number of tokens that match the vocabulary words
on a single-word basis, tokenization currently limits the vocabulary match encoding, and in
future research this must not impose a major limitation on being able to quantify the NLP
dimensions. It is also important that the dimension vocabularies have universally indicative
words and phrases and not only consist of examples found in news articles. Also related
to the NLP dimensions, there is also a limitation that some of the NLP dimension weights
may have to be adjusted in the future, for example potentially adding more weight to the
clickbait article title dimension and decreasing the weights of other dimensions. Although
it was concluded that the vast majority of articles did not have a clickbait title and its
quantified distribution did not have a large spread nor impact, an article title can still be
a significant indicator of whether an article is likely to be trustworthy or not. Additionally,
since the distribution of the normalized dimension count trustworthiness scores followed a
normal distribution, there were many articles that had a trustworthiness score between 45
and 55, meaning that there were many articles that were classified as likely trustworthy or
untrustworthy based on a few-point difference, therefore the NLP dimension weights may
have to be adjusted more in the future to fit realistic expectations.

There are also limitations regarding the overall functionality of the data tool. First, the
framework for quantifying the NLP dimensions is mandatory for the LLM and data science
tool to function properly and predict accurately. Without the inclusion of this framework,

104
the accuracy for the LLM predictions would be approximately 0.53 (the previous model
trainings without the NLP dimensions added produced correlation coefficients of around
0.73). The trustworthiness score predictions would be highly inaccurate especially due to
the predictions solely being reliable on the short truncated texts passed through the model,
thereby not enough context to assist with predictions, and this would yet remind that LLMs
have a limited token input when it comes to evaluating long texts, which fortunately this
thesis work was able to circumvent that limitation. Next, the data tool only requires the
fields of article text and article title in order to make predictions. The ultimate objective
of this thesis component was to devise a tool that could evaluate misinformation potential
solely on text and to require as few fields as possible. Additional fields including article
author, article publication, and other outside factors could potentially be difficult for users
of the tool to collect, however it should be noted that these factors can also make an impact
on whether a news article is likely to be trustworthy or not. Another limitation is that
this tool does not work using non-English-language news articles, as dimension vocabularies
contained only English words and punctuation. It is important to note that misinformation
detection in non-English languages is still a major research gap in the misinformation domain
[72], however this work will not be able to contribute to this research gap as compiling NLP
dimension vocabularies in non-English languages would require extensive time and effort and
likely a translation API that would add an additional monetary cost to research.

Another important limitation to note about Component 1 is that the base LLM was trained
on only around 50,000 news articles. Although these news articles were from a wide diversity
of sources, it is possible that the training dataset may have to be expanded to include 100,000
or even 200,000 news articles to ensure consistent predictions. In addition, as the news
articles in the training dataset become less recent, more recent news articles will have to be
added to include current information or instances of misinformation that have been verified.
The same goes for the NLP dimension vocabularies; these would also have to be updated
with current events.

105
A limitation to note about Component 3 is that there can be potential to ask a more
directed prompt to the ChatGPT and PaLM APIs to produce more explainable and accurate
trustworthiness scores. Despite specifically asking the APIs to return a trustworthiness score
between 0 and 100 (setting 0 as the lower bound and 100 as the higher bound) with the
attached article text, there were some drawbacks to the responses returned that can be
improved upon using prompt engineering. ChatGPT-3.5 overall had good explainability and
was able to interpret the article content fairly well, however at times it would not be able to
provide trustworthiness scores for some articles due to not enough context. Despite testing
these articles multiple times on the API to see if these responses would improve, they did not,
therefore these articles automatically had to be given a score of 0. Assigning these texts a 0
meant an inaccurate reflection of the true trustworthiness context in these articles and also
a loss of trustworthiness information that otherwise would have been valuable for analysis.
ChatGPT-4 for the majority of the time did not provide explainability for why it provided
trustworthiness scores for the news articles. Since reasons for scoring was an important
aspect of generative AI models to evaluate in misinformation detection, this inability to
provide that information left out important insights into what factors ChatGPT-4 could have
considered, especially if there were certain factors that ChatGPT-4 might have prioritized for
signifying trustworthiness compared to other factors that ChatGPT-3.5 or PaLM might have
considered. Finally, since PaLM tended to misinterpret the article context and/or origin,
prompts to produce better explainability need to be improved greatly.

Finally, there are a couple of minor limitations that need to be addressed. One is that since
the association rule mining was unable to be performed on the news article dataset but was
able to be performed on a true and false statement dataset, it is important to conclude
that association rule mining should only be used on textual data that is short in length and
it should not be recommended for use on datasets with long article texts. There are still
many opportunities to use association rule mining, using verified true and false statements
as well as social media data. Another limitation is that simple machine learning models (i.e.,

106
logistic regression, support vector machine, XGBoost) and deep learning models like neural
networks were not tested during the course of this thesis. It is possible that future research
could involve these models to see if they perform as better or as consistently, however these
models do not capture textual context of data unlike LLMs, but the inclusion of quantified
NLP dimensions could help the performances of these models.

6.3 Potential for Improvement

The NLP dimension vocabularies is a limitation that can be greatly improved on. First,
phrases will need to be added to the vocabularies to specify examples of misinformation that
could otherwise be misinterpreted if only spanning a single word (i.e., "steal the election"
versus "steal"). NLTK tokenization should be eliminated and instead string mentions should
be used for counting the instances of dimension vocabulary matches. Lemmatization, or an
NLP technique that reduces a verb to its basic form, could also be used to match lemmatized
versions of vocabulary phrases to reduce the effort of vocabulary compilation. Inclusion of
vocabularies that are not indicative of the dimensions could also be added to see if the
news articles contain more words indicating negative sentiment versus neutral sentiment,
more words indicating objective views versus biased views, etc. Improving the complexity
of the dimension vocabularies would also assist with improving dimension weights in the
future. The biggest limitation to improving the vocabularies however is that expanding
these to cover a wide variety of indicative words and phrases can take a lot of time and will
have to consistently be updated as more common instances of misinformation appear over
time. However, expanding the dimension vocabularies would greatly improve the quality
and reliability of this thesis work.

Likewise, since the distribution of the labeled scores followed a normal shape, making the
distribution follow a more bimodal shape through adjusting the dimension weights could help
classification tasks be more accurate in the future, therefore there is more work that can be

107
done to re-evaluate the proper weights of the NLP dimensions. While not all news articles
contain all 13 dimensions proposed in this thesis, and some were weighted less due to these
not being present all the time in news articles, cumulative scores could still be calculated by
determining whether a news article has at least a certain number of dimension requirements
met based on its context, and if so that article is trustworthy, otherwise it is not likely to be
trustworthy.

Next, the base LLM in the data tool can be trained on more data, even if the training
time will very likely take longer. Additional news articles from FakeNewsCorpus and other
open-source datasets would be very helpful to include even more variety in article origins,
and potentially non-English-language news articles could be incorporated as well to help
close the gap on non-English-language misinformation detection. However, compiling NLP
dimension vocabularies in non-English languages will involve a lot of effort to compile as
well.

Finally, improvements on prompt engineering related to Component 3 is another area to

improve upon and do future work in. For ChatGPT-3.5, prompts clarifying that for the
given text, even if there was not enough context to make it sound like a typical news article,
requiring the model to still evaluate the text for trustworthiness as it is would probably lead
to the generation of more scores. Results using ChatGPT-4 could likely improve by having
prompts adding the requirement that the model should provide reasons why the scores were
given. To improve explainability in PaLM’s responses, there should be clarifications that
require the model to evaluate the article text as it is, to assume that there is no author
nor article origin, and not to assume any common false claims were made unless specifically
seen in the text. There should also be a clarification to not give generic responses on
what factors should be considered to determine trustworthiness; it is more important to ask
the model itself how it would rate this article text and why from its perspective. For all
3 models, including context related to the interdisciplinary fields of communications and

108
psychology in the prompts can lead to more fine-tuned results (i.e., "please consider factors
such as psychological variables that can influence perception of the article content as well
as communication variables including sentiment, bias, logical coherence, and any other text-
communication factors that are visible in the text"). Although some studies have been done
using generative AI models for misinformation detection, future work could include more
studies under research contexts, and possibly involving evaluation of the new Google Gemini
API (as of February 2024, the PaLM API has been deprecated). To recommend that these
models are reliable for open-source misinformation detection tasks, an accuracy of at least
80-90% would be needed as if these models were to perform that well, it would make them
likely more trustworthy for public users in terms of performance.

6.4 Societal Implications

In Chapter 1 of this thesis, there were some target stakeholders that were introduced:

• U.S. Government agencies

• U.S. Military agencies

• Industry programmers

• Public users

The development of the data science tool through this study can definitely be used by the
target stakeholders for rapid automated trustworthiness scoring, however it may take signifi-
cant improvements to the framework in order for the tool to be validated for open-source and
government use. The target stakeholders can greatly benefit from this thesis work as another
data tool can be used for the specified purposes of predicting comprehensive trustworthiness
scores for a wide variety of news articles, with added transparency to improve the trust of the
technology with users. This tool could likewise be potentially instrumental in information
warfare by government agencies that may currently be investing in tactics against foreign

109
and domestic adversaries that spread harmful misinformation. Military agencies can also use
this tool to combat adversaries that may spread defaming misinformation about potential
suppliers that may be crucial to supply chain management for the U.S. military; having a
truthful image about which suppliers are trustworthy to meet supply needs for the military
is important to keep supply chain operations and military readiness strong. For industry
programmers, the tool can also help these programmers rapidly and accurately populate
trustworthiness scores for databases of news articles, helping facilitate misinformation de-
tection tasks in industry.

Another impact to society is that this study can yield a warning to public users who may be
considering using AI models or APIs to determine whether news they see on the internet,
mainstream media, etc. is trustworthy or not. Based on the results of the thesis, it would
not be recommended that open-source users rely on ChatGPT or Bard (the user interface
version of PaLM) to prompt if certain news articles are misinformation due to their inaccu-
racies and potentially misleading explanations in their responses. Unless these AI models are
further developed to be more reliable and transparent, users should not use these platforms
for misinformation detection.

6.5 Summary of Limitations and Improvements

Again, here is a tabular summary of the limitations of this thesis along with suggested
improvements to these limitations:

110
 Figure 6.1: Summary of Limitations and Potential Improvements

6.6 Note on Subjectivity of Misinformation

Different individuals can have differing opinions on what news is considered misinformation
or not. While it is very important for humans in the loop to analyze misinformation from
an objective lens, some people may tend to classify news articles as likely not to be mis-

111
information that may be, for example, noticeably higher in sentiment or bias. Potentially,
some people who believe in conspiracy theories as truth may label a training dataset of news
articles much differently compared to people who would try to evaluate the trustworthiness
of news articles with as little personal opinion as possible. Thus, different interpretations of
what constitutes misinformation, based on individuals’ subjective views of information, can
impact whether these individuals would be able to find an automated tool trustworthy to
use due to personal disagreements on what the tool might classify as misinformation or not.

112
Chapter 7

Conclusion

7.1 Summary

This thesis work has yielded significant findings that can be of considerable value to the
field of data science and to society itself. Multiple types of models were tested to determine
which one could be most practical for misinformation detection research, investment in in-
formation warfare, and public-use misinformation identification in the future, and based on
the results comparing these models limitations and strengths in each were discovered. In
addition, an explainable data tool was proposed that could generate rapid, accurate trust-
worthiness scores for news articles based on communication dimensions of misinformation
present in those texts while avoiding the limitation of the maximum input token length for
the base LLM. Finally, results from evaluating some AI models suggested that these models
have significant disadvantages, and have a serious potential to be unreliable, in detecting
misinformation accurately.

The most crucial limitation of this work was that the vocabulary for encoding the commu-
nication dimensions of misinformation needs to be upgraded over time. Other limitations
included that the training dataset may likely have to be expanded and that the inclusion
of the NLP dimensions are likewise required in order for the proposed data tool to function
properly and accurately. Improvements to help close research gaps in non-English misinfor-
mation detection may also be utilized in the data tool in the future. Simple machine learning

113
models and neural networks were left out of the comparison analysis between LLMs, unsu-
pervised learning, and AI models, which while fine-tuned LLMs proved to be very accurate,
it is still important to take all supervised learning models into consideration to determine if
these simple machine learning models can still be used for misinformation detection in the
future. Finally, the unsupervised learning method of association rule mining was unable to
be performed on the news article dataset, yet this technique can still have the potential to
reveal valuable insights about verified true and false statements, and some work involving
this experimentation was able to be completed as supplementary work.

7.2 Contribution to Data Science Field

To again summarize the significance of the findings of this thesis to general knowledge, here
is what each of the components accomplished.

Component 1 (Data Tool with LLM):

• Added a data science tool option to combat misinformation.

• Created a tool that can evaluate news articles for trustworthiness.

• Verified that the LLM is still a solidly reliable type of model for misinformation detec-
tion.

• Contributed explainable machine learning for misinformation detection.

• Bypassed the maximum input token length for the base LLM through NLP dimensions.

• Contributed universal dimensions of misinformation that should be considered in news

article content.

Component 2 (Unsupervised Learning):

• Generated some practical data insights that differentiated trustworthy and untrust-
worthy news articles through anomaly detection and LDA topic modeling.

114
 • Could not verify cosine similarity and t-SNE as an effective method for misinformation
detection.

• Tentatively concluded that unsupervised learning methods could still be used for mis-
information detection research and can make contributions to automated fact-checking
research.

• Contributed unique unsupervised learning methods as this type of model is not com-
monly used for misinformation detection.

Component 3 (Generative AI Model Evaluation):

• Concluded that generative AI models are generally less accurate compared to the other
types of models tested in this thesis.

• Could not conclude that AI models would be recommended for misinformation detec-
tion as these models had an average classification accuracy of 64%.

• Also concluded that AI models provide faulty explainability for trustworthiness scores.

• Suggested that AI models should not be used by public users for misinformation de-
tection but can be used for research purposes for further studies.

To synthesize, all 3 components of this thesis helped provide a better understanding of the
strengths and limitations of each type of model used for misinformation detection.

7.3 Contribution to Society and Stakeholders

The significance of the findings of the thesis to society and the intended stakeholders provide
the impacts described below.

Society:

115
 • Provided a big data solution to help protect society against the harmful effects of
misinformation through more accurate trustworthiness scoring of news articles.

• Proposed an data tool that can determine how trustworthy a given news article is,
whether originated from mainstream, independent, or social media news outlets.

• Included explainability to build trust in the tool with the user.

• Warned that AI models should not be depended upon by public users for misinfor-
mation detection as responses can tend to be unreliable, which could impact people’s
perspectives on popular generative AI models.

Stakeholders:

• Government agencies: an added data tool to help invest in combating misinforma-

tion spread through harmful fake news.

• Military agencies: a data tool that can also combat misinformation that can impact
business decisions, which in turn can disrupt crucial supply chain operations for the
military.

• Industry programmers: a tool that can rapidly populate accurate trustworthiness

scores for databases of news article texts with explainability metrics.

• Public users: an open-source tool that can be used to holistically determine if news
found on a diversity of news outlets is trustworthy or not.

7.4 Future Work and Recommendations

Future work will likely require expanding the dimension vocabularies to include recent veri-
fied instances of misinformation including specified phrases, words, and punctuation. How-
ever, compiling these extensive vocabularies would require a lot of time and research in

116
order to not omit important indicative features. Yet, this improvement to the thesis work
would help immensely with the quality of the finished product. Other work would also
include training the base LLM on more news articles and potentially including non-English-
language articles and dimension vocabularies. After finalizing the vocabularies, peer review
is welcome to evaluate and certify the tool to be ready for open-source and government use.
There can also be work done to adapt the tool to accommodate social media data; many
features may have to be altered, but social media misinformation’s impacts are just as pro-
found as news misinformation’s. Therefore, having a similar tool for social media would be
just as practical for analysis and comprehensive trustworthiness predictions.

Regarding the limitation that association rule mining raised, future work can also include
using association rule mining and the unsupervised learning methods used in this thesis to
differentiate verified true and false statements. As mentioned at the end of Chapter 4, work
has already been completed on this, however this work was not able to be included in the
thesis, with the exception of some data visualizations that are shown in Chapter 10, because
this work was done on a completely different dataset. Thereby it would digress from the
conclusions drawn from comparing the LLMs, unsupervised learning methods, and AI mod-
els being performed on the news article dataset. This work would still help contribute to
automated fact-checking research, which is equally instrumental in field research on misin-
formation.

As for recommendations for future research, there can be several encouragements given the
promising results of this thesis. First, despite the insufficient results that the AI models pro-
vided from the evaluation and some previous studies from the literature review, since these
models are fairly recent there still needs to be additional comprehensive studies on their
ability to detect misinformation conducted, both in accuracy and in explainability. The
prompts for the generative AI models can likewise be improved through prompt engineering
to produce more trustworthiness scores and explainability for given article texts so as to not

117
lead to a loss of information in analyzing these models’ efficacy in misinformation detection.
While adjustments through adding clarifications to the prompts can produce more results, it
should be noted that the AI models may still not be able to capture all context between the
prompt and article text combined if there is too much complexity, leading to responses that
may not be able to answer all aspects of the prompt. Therefore, keeping prompts relatively
concise while adding enough specified context to make the prompts as effective as possible
will be key to producing better results.

Next, since LLMs have been consistently accurate in determining trustworthiness scores for
news articles, other supervised learning models should likewise be evaluated (i.e., simple
machine learning models, neural networks) in their ability to detect misinformation. If these
models also stand, they can likewise be recommended for viable misinformation detection
research. This can also apply to other unsupervised learning models. Recommended work
would also include re-evaluating the NLP dimension weights and vocabularies as well as
conducting more studies on generative AI models’ use for misinformation detection. In the
future, misinformation detection research should incorporate communication dimensions to
improve accuracy, as this would be an effective intersection between the fields of data science
and communications to provide an interdisciplinary solution to the major societal problem
of misinformation.

7.5 Final Thoughts

The main takeaway from this thesis is that given the vast amount of research done, and
research gaps, on misinformation detection methods, it is important to resolve which meth-
ods, models, and/or tools are most effective at completing the crucial task of differentiating
which is misinformation and which is not. The news is an important part of free speech and
journalism in society, however misinformation is an abuse of free speech and can endanger
reputations of citizens, organizations, and countries by spreading misleading information to

118
its readers. The ideal solution to the problem of misinformation is widespread media literacy
and legislative regulations, however current solutions have required involvement in the data
science field which has led to the creation of tools and models to combat misinformation.
Target stakeholders likewise need to have awareness that such tools exist in order to be
educated on which news is true or fake, with technological accountability to establish trust
in these tools. Despite the harrowing task of winning the information war, data science has
a major role to play in winning this war and to ensure the trustworthiness and effectiveness
of tools past, present, and future.

119
Chapter 8

Acknowledgements

I would like to express my deepest gratitude to the following individuals and organizations
who have played a significant role in the completion of this master’s thesis.

First, I am very grateful to my thesis committee chair and advisor, Dr. David Ebert, for
providing me with the utmost support throughout this research process. I would also like to
extend my sincere appreciation to Dr. Gopi Danala and Dr. Wolfgang Jentner for providing
crucial advising and feedback for this thesis.

Second, I would like to thank the members of my thesis committee. I greatly appreciate
Dr. Dean Hougen and Dr. Naveen Kumar for their support and for helping shape the
data-driven narrative and components of this research. As misinformation is not just a data
science issue and is likewise closely related to the fields of communications and journalism, I
am deeply grateful for also having Dr. Jeong-Nam Kim and Dr. Katerina Tsetsura from the
Gaylord College of Journalism and Mass Communication on the committee. Their expertise
and steadfast support have really helped in providing quality, transdisciplinary work for this
thesis.

Finally and most importantly, I would like to thank the Data Institute for Societal Chal-
lenges (DISC) here at the University of Oklahoma. I am deeply grateful to DISC for helping
financially support this research and for providing me with a graduate research assistantship
that has become invaluably crucial to my research experience and development as a data

120
scientist. This assistantship has also relieved many financial burdens associated with pursu-
ing a master’s degree, and I am profoundly thankful for that as without this support, this
thesis would not have been possible.

121
Chapter 9

References

[1] Adel, H., Dahou, A., Mabrouk, A., Abd Elaziz, M., Kayed, M., El-Henawy, I. M., ...
& Amin Ali, A. (2022). Improving crisis events detection using distilbert with hunger
games search algorithm. Mathematics, 10 (3), 447.

[2] Akhtar, P., Ghouri, A. M., Khan, H. U. R., Amin ul Haq, M., Awan, U., Zahoor, N.,
... & Ashraf, A. (2023). Detecting fake news and disinformation using artificial intel-
ligence and machine learning to avoid supply chain disruptions. Annals of Operations
Research, 327 (2), 633-657.

[3] Amer, E., Kwak, K. S., & El-Sappagh, S. (2022). Context-based fake news detection
model relying on deep learning models. Electronics, 11 (8), 1255.

[4] Azzimonti, M., & Fernandes, M. (2023). Social media networks, fake news, and polar-
ization. European journal of political economy, 76, 102256.

[5] Bhat, M. M. (2022). Study of Effectiveness of Stylometry in Misinformation Detection

(Doctoral dissertation, The Ohio State University).

[6] Calvillo, D. P., Garcia, R. J., Bertrand, K., & Mayers, T. A. (2021). Personality
factors and self-reported political news consumption predict susceptibility to political
fake news. Personality and individual differences, 174, 110666.

122
 [7] Caramancion, K. M. (2023). News Verifiers Showdown: A Comparative Performance
Evaluation of ChatGPT 3.5, ChatGPT 4.0, Bing AI, and Bard in News Fact-Checking.
arXiv preprint arXiv:2306.17176.

[8] Caruana, R., & Niculescu-Mizil, A. (2006, June). An empirical comparison of su-
pervised learning algorithms. In Proceedings of the 23rd international conference on
Machine learning (pp. 161-168).

[9] Chandola, V., Banerjee, A., & Kumar, V. (2009). Anomaly detection: A survey. ACM
computing surveys (CSUR), 41 (3), 1-58.

[10] Chiu, M. M., Morakhovski, A., Ebert, D., Reinert, A., & Snyder, L. S. (2023). De-
tecting COVID-19 fake news on Twitter: Followers, emotions, relationships, and un-
certainty. American Behavioral Scientist, 00027642231174329.

[11] Chiu, M. M., & Oh, Y. W. (2021). How fake news differs from personal lies. American
behavioral scientist, 65 (2), 243-258.

[12] Chiu, M. M., Park, C. H., Lee, H., Oh, Y. W., & Kim, J. N. (2022). Election Fraud
and Misinformation on Twitter: Author, Cluster, and Message Antecedents. Media
and Communication, 10 (2), 66-80.

[13] Cui, Z. (2022, January). COVID-19 Fake News and Misinformation Detection using
Transformer Learning. In 2022 3rd International Conference on Education, Knowledge
and Information Management (ICEKIM) (pp. 965-968).

[14] Cunha, B., & Manikonda, L. (2022). Classification of Misinformation in New Articles
using Natural Language Processing and a Recurrent Neural Network. arXiv preprint
arXiv:2210.13534.

[15] Dame Adjin-Tettey, T. (2022). Combating fake news, disinformation, and misinforma-
tion: Experimental evidence for media literacy education. Cogent arts & humanities,

123
 9 (1), 2037229.

[16] De Magistris, G., Russo, S., Roma, P., Starczewski, J. T., & Napoli, C. (2022). An
explainable fake news detector based on named entity recognition and stance classifi-
cation applied to covid-19. Information, 13 (3), 137.

[17] Deng, J., & Lin, Y. (2022). The benefits and challenges of ChatGPT: An overview.
Frontiers in Computing and Intelligent Systems, 2 (2), 81-83.

[18] Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). Bert: Pre-training of deep
bidirectional transformers for language understanding. arXiv preprint
arXiv:1810.04805.

[19] Duan, Y., Edwards, J. S., & Dwivedi, Y. K. (2019). Artificial intelligence for decision
making in the era of Big Data–evolution, challenges and research agenda. International
journal of information management, 48, 63-71.

[20] Fernando, A., & Wijayasiriwardhane, T. K. (2020, September). Identifying religious

extremism-based threats in SriLanka using bilingual social media intelligence. In
2020 International Research Conference on Smart Computing and Systems Engineering
(SCSE) (pp. 103-110). IEEE.

[21] Gradoń, K. T., Hołyst, J. A., Moy, W. R., Sienkiewicz, J., & Suchecki, K. (2021).
Countering misinformation: A multidisciplinary approach. Big Data & Society, 8 (1),
20539517211013848.

[22] Grignolio, A., Morelli, M., & Tamietto, M. (2022). Why is fake news so fascinating to
the brain?. European Journal of Neuroscience, 56 (11), 5967-5971.

[23] Guess, A., Nyhan, B., & Reifler, J. (2018). Selective exposure to misinformation:
Evidence from the consumption of fake news during the 2016 US presidential campaign.

124
[24] Guimarães, N., Figueira, Á., & Torgo, L. (2021). Can fake news detection models main-
tain the performance through time? A longitudinal evaluation of twitter publications.
Mathematics, 9 (22), 2988.

[25] Gunel, B., Du, J., Conneau, A., & Stoyanov, V. (2020). Supervised contrastive learning
for pre-trained language model fine-tuning. arXiv preprint arXiv:2011.01403.

[26] Hameleers, M., & Brosius, A. (2022). You are wrong because I am right! The perceived
causes and ideological biases of misinformation beliefs. International Journal of Public
Opinion Research, 34 (1), edab028.

[27] Haupt, M. R., Li, J., & Mackey, T. K. (2021). Identifying and characterizing scientific
authority-related misinformation discourse about hydroxychloroquine on twitter using
unsupervised machine learning. Big Data & Society, 8 (1), 20539517211013843.

[28] Hoes, E., Altay, S., & Bermeo, J. (2023). Using ChatGPT to Fight Misinformation:
ChatGPT Nails 72% of 12,000 Verified Claims.

[29] Islam, M. R., Liu, S., Wang, X., & Xu, G. (2020). Deep learning for misinformation
detection on online social networks: a survey and new perspectives. Social Network
Analysis and Mining, 10, 1-20.

[30] Jamalzadeh, S., Barker, K., González, A. D., & Radhakrishnan, S. (2022). Protect-
ing infrastructure performance from disinformation attacks. Scientific Reports, 12 (1),
12707.

[31] Jang, S. M., & Kim, J. K. (2018). Third person effects of fake news: Fake news
regulation and media literacy interventions. Computers in human behavior, 80, 295-
302.

[32] Jelodar, H., Wang, Y., Yuan, C., Feng, X., Jiang, X., Li, Y., & Zhao, L. (2019).
Latent Dirichlet allocation (LDA) and topic modeling: models, applications, a survey.

125
 Multimedia Tools and Applications, 78, 15169-15211.

[33] Jwa, H., Oh, D., Park, K., Kang, J. M., & Lim, H. (2019). exbake: Automatic fake
news detection model based on bidirectional encoder representations from transformers
(bert). Applied Sciences, 9 (19), 4062.

[34] Kertysova, K. (2018). Artificial intelligence and disinformation: How AI changes the
way disinformation is produced, disseminated, and can be countered. Security and
Human Rights, 29 (1-4), 55-81.

[35] Khanum, M., Mahboob, T., Imtiaz, W., Ghafoor, H. A., & Sehar, R. (2015). A
survey on unsupervised machine learning algorithms for automation, classification and
maintenance. International Journal of Computer Applications, 119 (13).

[36] Khanzode, K. C. A., & Sarode, R. D. (2020). Advantages and disadvantages of artificial
intelligence and machine learning: A literature review. International Journal of Library
& Information Science (IJLIS), 9 (1), 3.

[37] Kochhar, S. K., & Kaur, R. (2023). Breaking the Taboos: Deploying Knowledge
Differentiation to study COVID-19 ramifications on Women’s Menstrual Health. In-
ternational Journal of Computing and Digital Systems, 13 (1), 1-1.

[38] Kong, S. H., Tan, L. M., Gan, K. H., & Samsudin, N. H. (2020, April). Fake news
detection using deep learning. In 2020 IEEE 10th symposium on computer applications
& industrial electronics (ISCAIE) (pp. 102-107). IEEE.

[39] Kumbhare, T. A., & Chobe, S. V. (2014). An overview of association rule mining
algorithms. International Journal of Computer Science and Information Technologies,
5 (1), 927-930.

[40] Kwek, A., Peh, L., Tan, J., & Lee, J. X. (2023). Distractions, analytical thinking
and falling for fake news: A survey of psychological factors. Humanities and Social

126
 Sciences Communications, 10 (1), 1-12.

[41] Liu, F. T., Ting, K. M., & Zhou, Z. H. (2008, December). Isolation forest. In 2008
eighth ieee international conference on data mining (pp. 413-422). IEEE.

[42] Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., ... & Stoyanov, V. (2019).
Roberta: A robustly optimized bert pretraining approach. arXiv preprint
arXiv:1907.11692.

[43] Martel, C., Pennycook, G., & Rand, D. G. (2020). Reliance on emotion promotes
belief in fake news. Cognitive research: principles and implications, 5, 1-20.

[44] McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in
nervous activity. The bulletin of mathematical biophysics, 5, 115-133.

[45] Meyer, J. G., Urbanowicz, R. J., Martin, P. C., O’Connor, K., Li, R., Peng, P. C., ... &
Moore, J. H. (2023). ChatGPT and large language models in academia: opportunities
and challenges. BioData Mining, 16 (1), 20.

[46] Mhatre, S., & Masurkar, A. (2021, June). A hybrid method for fake news detection
using cosine similarity scores. In 2021 International Conference on Communication
information and Computing Technology (ICCICT) (pp. 1-6). IEEE.

[47] Min, E., Rong, Y., Bian, Y., Xu, T., Zhao, P., Huang, J., & Ananiadou, S. (2022).
Divide-and-Conquer: Post-User Interaction Network for Fake News Detection on Social
Media. In Proceedings of the ACM Web Conference 2022 (pp. 1148-1158).

[48] Ng, L. H., & Taeihagh, A. (2021). How does fake news spread? Understanding path-
ways of disinformation spread through APIs. Policy & Internet, 13 (4), 560-585.

[49] Oh, Y. W., & Park, C. H. (2021). Machine cleaning of online opinion spam: Developing
a machine-learning algorithm for detecting deceptive comments. American behavioral
scientist, 65 (2), 389-403.

127
[50] Oliva, C., Palacio-Marín, I., Lago-Fernández, L. F., & Arroyo, D. (2022, August). Ru-
mor and clickbait detection by combining information divergence measures and deep
learning techniques. In Proceedings of the 17th International Conference on Availabil-
ity, Reliability and Security (pp. 1-6).

[51] Paredes, C. M. G., Machuca, C., & Claudio, Y. M. S. (2023). ChatGPT API: Brief
overview and integration in Software Development. International Journal of Engineer-
ing Insights, 1 (1), 25-29.

[52] Przybyła, P., & Soto, A. J. (2021). When classification accuracy is not enough: Ex-
plaining news credibility assessment. Information Processing & Management, 58 (5),
102653.

[53] Rahutomo, F., Kitasuka, T., & Aritsugi, M. (2012, October). Semantic cosine simi-
larity. In The 7th international student conference on advanced science and technology
ICAST (Vol. 4, No. 1, p. 1).

[54] Rai, N., Kumar, D., Kaushik, N., Raj, C., & Ali, A. (2022). Fake News Classifica-
tion using transformer based enhanced LSTM and BERT. International Journal of
Cognitive Computing in Engineering, 3, 98-105.

[55] Ramadasa, I., Liyanage, L., Asanka, D., & Dilanka, T. (2022). Analysis of the effec-
tiveness of using google translations api for nlp of sinhalese.

[56] Ray, P. P. (2023). ChatGPT: A comprehensive review on background, applications,

key challenges, bias, ethics, limitations and future scope. Internet of Things and Cyber-
Physical Systems.

[57] Reis, J. C., Correia, A., Murai, F., Veloso, A., & Benevenuto, F. (2019). Supervised
learning for fake news detection. IEEE Intelligent Systems, 34 (2), 76-81.

128
[58] Riego, N. C. R., & Villarba, D. B. (2023). Utilization of Multinomial Naive Bayes
Algorithm and Term Frequency Inverse Document Frequency (TF-IDF Vectorizer) in
Checking the Credibility of News Tweet in the Philippines. arXiv preprint
arXiv:2306.00018.

[59] Sanh, V., Debut, L., Chaumond, J., & Wolf, T. (2019). DistilBERT, a distilled version
of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108.

[60] Serrano, J. C. M., Papakyriakopoulos, O., & Hegelich, S. (2020, July). NLP-based
feature extraction for the detection of COVID-19 misinformation videos on YouTube.
In Proceedings of the 1st Workshop on NLP for COVID-19 at ACL 2020.

[61] Seth, N. (2021, August 26). Part 2: Topic modeling and Latent Dirichlet allocation
(LDA) using Gensim and Sklearn. Analytics Vidhya. https://www.analyticsvidhya.
com/blog/2021/06/part-2-topic-modeling-and-latent-dirichlet-allocation-lda-using-
gensim-and-sklearn/

[62] Shahi, G. K., & Nandini, D. (2020). FakeCovid–A multilingual cross-domain fact check
news dataset for COVID-19. arXiv preprint arXiv:2006.11343.

[63] Sharma, D. K., Garg, S., & Shrivastava, P. (2021, February). Evaluation of tools and
extension for fake news detection. In 2021 International Conference on Innovative
Practices in Technology and Management (ICIPTM) (pp. 227-232).

[64] Shinde, A. (2022, May). Unsupervised Detection of Misinformation in Financial State-

ments. In The International FLAIRS Conference Proceedings (Vol. 35).

[65] Shu, K., Mahudeswaran, D., & Liu, H. (2019). FakeNewsTracker: a tool for fake news
collection, detection, and visualization. Computational and Mathematical Organization
Theory, 25, 60-71.

129
[66] Su, Q., Wan, M., Liu, X., & Huang, C. R. (2020). Motivations, methods and metrics of
misinformation detection: an NLP perspective. Natural Language Processing Research,
1 (1-2), 1-13.

[67] Tamkin, A., Brundage, M., Clark, J., & Ganguli, D. (2021). Understanding the ca-
pabilities, limitations, and societal impact of large language models. arXiv preprint
arXiv:2102.02503.

[68] Thorne, J., Vlachos, A., Christodoulopoulos, C., & Mittal, A. (2018). FEVER: a large-
scale dataset for fact extraction and VERification. arXiv preprint arXiv:1803.05355.

[69] Titiliuc, C., Ruseti, S., & Dascalu, M. (2020, September). What’s Been Happening
in the Romanian News Landscape? A Detailed Analysis Grounded in Natural Lan-
guage Processing Techniques. In 2020 22nd International Symposium on Symbolic and
Numeric Algorithms for Scientific Computing (SYNASC) (pp. 195-201). IEEE.

[70] Tucker, Joshua A., et al. "Social media, political polarization, and political disin-
formation: A review of the scientific literature." Political polarization, and political
disinformation: a review of the scientific literature (March 19, 2018) (2018).

[71] Turner, J., Kantardzic, M., Vickers-Smith, R., & Brown, A. G. (2023). Detecting
Tweets Containing Cannabidiol-Related COVID-19 Misinformation Using Transformer
Language Models and Warning Letters From Food and Drug Administration: Content
Analysis and Identification. JMIR infodemiology, 3 (1), e38390.

[72] van de Meerakker, K. O. (2022). The foreign language effect on the credibility of fake
news messages among Dutch news readers and the influence of emotional loading.

[73] Van der Maaten, L., & Hinton, G. (2008). Visualizing data using t-SNE. Journal of
machine learning research, 9 (11).

130
[74] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... &
Polosukhin, I. (2017). Attention is all you need. Advances in neural information
processing systems, 30.

[75] Verma, P. K., Agrawal, P., Amorim, I., & Prodan, R. (2021). WELFake: word embed-
ding over linguistic features for fake news detection. IEEE Transactions on Computa-
tional Social Systems, 8 (4), 881-893.

[76] Wadud, M. A. H., Mridha, M. F., & Rahman, M. M. (2022). Word embedding methods
for word representation in deep learning for natural language processing. Iraqi Journal
of Science, 1349-1361.

[77] Waisberg, E., Ong, J., Masalkhi, M., Zaman, N., Sarker, P., Lee, A. G., & Tavakkoli,
A. (2023). Google’s AI chatbot “Bard”: a side-by-side comparison with ChatGPT and
its utilization in ophthalmology. Eye, 1-4.

[78] Wang, W. Y. (2017). " liar, liar pants on fire": A new benchmark dataset for fake
news detection. arXiv preprint arXiv:1705.00648.

[79] Watts, D. J., Rothschild, D. M., & Mobius, M. (2021). Measuring the news and
its impact on democracy. Proceedings of the National Academy of Sciences, 118 (15),
e1912443118.

[80] Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., ... & Rush, A.
M. (2019). Huggingface’s transformers: State-of-the-art natural language processing.
arXiv preprint arXiv:1910.03771.

[81] Zhang, C., Gupta, A., Kauten, C., Deokar, A. V., & Qin, X. (2019). Detecting fake
news for reducing misinformation risks using analytics approaches. European Journal
of Operational Research, 279 (3), 1036-1052.

131
[82] Zhang, X. Examining COVID-19 Vaccination Misinformation and Clarification by the
Public Sector in Hong Kong.

[83] Zhong, H., & Mei, H. (2017). An empirical study on API usages. IEEE Transactions
on Software Engineering, 45(4), 319-334.

132
Chapter 10

Appendices

10.1 Appendix A: Supplementary Code Snippets

10.1.1 NLP Dimension Vocabularies

Figure 10.1: Sentiment Dimension Vocabulary

Figure 10.2: Persuasion Dimension Vocabulary

133
 Figure 10.3: Exaggeration Dimension Vocabulary

Figure 10.4: Context Dimension Vocabulary

Figure 10.5: Multiple Perspectives Dimension Vocabulary

Figure 10.6: Statistics Dimension Vocabulary

134
 Figure 10.7: Previous Articles Dimension Vocabulary

Figure 10.8: Distraction Dimension Vocabulary

Figure 10.9: Claim Verification Dimension Vocabulary

Figure 10.10: Logical Coherence Dimension Vocabulary

135
 Figure 10.11: Clickbait Title Dimension Vocabulary

10.1.2 Data Tool Functions

Figure 10.12: Function to Count and Normalize Vocabulary

136
 Figure 10.13: Function to Extract Named Entities

Figure 10.14: Function to Count and Normalize Named Entities

137
Figure 10.15: Function to Count and Normalize Statistics Count

Figure 10.16: Function to Normalize TF-IDF Vectors

138
 Figure 10.17: Function to Set NLP Dimension Scores (Truncated)

Figure 10.18: Function to Set NLP Dimension Scores (Continued, Truncated)

139
 Figure 10.19: Function to Concatenate Dimensions

Figure 10.20: Importing the RoBERTa Weights and Compiling the Model

Figure 10.21: Making Predictions Using the RoBERTa Model

140
10.2 Appendix B: Additional Visualizations and Images

10.2.1 LLM Training Versus Validation

Figure 10.22: BERT Training vs. Validation Loss

141
Figure 10.23: RoBERTa Training vs. Validation Loss

Figure 10.24: DistilBERT Training vs. Validation Loss

142
 Figure 10.25: BERT Training vs. Validation Mean Absolute Error

Figure 10.26: RoBERTa Training vs. Validation Mean Absolute Error

143
 Figure 10.27: DistilBERT Training vs. Validation Mean Absolute Error

10.2.2 Association Rule Mining on Politifact Kaggle Dataset

Figure 10.28: 2-D Scatter Plot of Politifact Kaggle Dataset Association Rules

144
Figure 10.29: 3-D Scatter Plot of Politifact Kaggle Dataset Association Rules

Figure 10.30: Network Plot of True Statement Association Rules

145
Figure 10.31: Network Plot of Pants-Fire Statement Association Rules

146
Figure 10.32: Heatmap of True Statement Association Rules

147
 Figure 10.33: Heatmap of Pants-Fire Association Rules

Figure 10.34: Sankey Diagram of True Statement Association Rules

148
 Figure 10.35: Sankey Diagram of Pants-Fire Statement Association Rules

10.3 Appendix C: Glossary of Commonly Used Terms

and Acronyms

Application Programming Interface (API): Unlike a user interface, a software in-

termediary that allows a user to communicate with and retrieve information from a web
application.

Artificial Intelligence (AI): Intelligent machines or software that can require training on
large amounts of data. Recent versions of generative AI like OpenAI’s ChatGPT can imitate
chatbots and generate new data based on previous training data.

Bidirectional Encoder Representations from Transformers (BERT): One of the first

LLMs developed to expand upon the transformer model framework and is an open-source
model used for language prediction and other self-supervised NLP tasks.

Disinformation: Misleading or false information with intent to deceive. In the context of

this thesis, opinionated news is an example of disinformation.

Distilled BERT (DistilBERT): An LLM that retains almost the same effectiveness as
BERT, however it is lighter to deploy and use.

Large Language Model (LLM): A language model trained on vast amounts of textual
data to perform general language generation and prediction tasks.

149
Latent Dirichlet Allocation (LDA): An unsupervised learning model that is used to
classify text in a document or corpus of text to particular topics.

Misinformation: Misleading or false information without intent to deceive.

Natural Language Processing (NLP): Computer programming that converts textual

data into machine interpretations of human language.

Pathways Language Model (PaLM): The model framework and API version of Google
Bard.

Robustly Optimized BERT Approach (RoBERTa): An LLM developed to outper-

form BERT through being trained on more data and using a more robust training approach
of dynamic masking.

t-Distributed Stochastic Neighbor Embedding (t-SNE): A type of unsupervised

learning model that can visualize high-dimensional data clusters using reduced dimensions.

Term Frequency-Inverse Document Frequency (TF-IDF): A measure of importance

of a word to a document in a corpus given that words can appear more frequently than
others. The TF-IDF formula in the form

 
N
wi,j = tfi,j × log
dfi

is used to calculate the weight wi,j of term i in document j, where:

tfi,j = frequency of term i in document j

dfi = number of documents containing term i

N = total number of documents in the corpus

Trustworthiness: A comprehensive measure of misinformation likelihood, of a text, based

150
on multiple communication factors. A higher trustworthiness means a lower likelihood of
misinformation, and vice versa.

Unsupervised Learning: A type of machine learning that does not require training of
a labeled dataset to make predictions. An example of this type of learning is generating
clusters of data points in a dataset.

151