Scribd Logo
Upload

Welcome to Scribd!

  • 11 views

    Projet Scientifique

    Uploaded by

    Cesar Ndjoko
    The document outlines a presentation on developing a fast model for sentiment analysis of social media. It discusses the problem of slow sentiment analysis models and the objectives to assess existing techniques, improve on deficiencies, and propose a new technique. The literature review covers lexicon-based and machine learning approaches from previous works. One work built an emotion lexicon through simple techniques to improve precision and coverage. Another used support vector machines, naive bayes, and maximum entropy classifiers in a voting approach to classify sentiment and measure depression from social media data.

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd

    Projet Scientifique

    Uploaded by

    Cesar Ndjoko
    11 views34 pages
    The document outlines a presentation on developing a fast model for sentiment analysis of social media. It discusses the problem of slow sentiment analysis models and the objectives to assess existing techniques, improve on deficiencies, and propose a new technique. The literature review covers lexicon-based and machine learning approaches from previous works. One work built an emotion lexicon through simple techniques to improve precision and coverage. Another used support vector machines, naive bayes, and maximum entropy classifiers in a voting approach to classify sentiment and measure depression from social media data.

    Original Title

    Projet scientifique.pptx

    Copyright

    © © All Rights Reserved

    Available Formats

    PPTX, PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    The document outlines a presentation on developing a fast model for sentiment analysis of social media. It discusses the problem of slow sentiment analysis models and the objectives to assess existing techniques, improve on deficiencies, and propose a new technique. The literature review covers lexicon-based and machine learning approaches from previous works. One work built an emotion lexicon through simple techniques to improve precision and coverage. Another used support vector machines, naive bayes, and maximum entropy classifiers in a voting approach to classify sentiment and measure depression from social media data.

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd
    Download as pptx, pdf, or txt
    11 views34 pages

    Projet Scientifique

    Uploaded by

    Cesar Ndjoko
    The document outlines a presentation on developing a fast model for sentiment analysis of social media. It discusses the problem of slow sentiment analysis models and the objectives to assess existing techniques, improve on deficiencies, and propose a new technique. The literature review covers lexicon-based and machine learning approaches from previous works. One work built an emotion lexicon through simple techniques to improve precision and coverage. Another used support vector machines, naive bayes, and maximum entropy classifiers in a voting approach to classify sentiment and measure depression from social media data.

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd
    Download as pptx, pdf, or txt
    You are on page 1of 34

    A FAST MODEL OF SENTIMENT ANALYSIS FOR

    SOCIAL MEDIA

    By NDJOKO A BIDIAS Cesar Junior FE21P043

    SUPERVISED BY

    Pr. FUTE ELIE Tuesday, 17 January 2022


    OUTLINE OF THE PRESENTATION

    ❏ Introduction
    ❏ Problem Statement
    ❏ Objectives
    ❏ Main Objective
    ❏ Specific Objectives
    ❏ Literature Review
    ❏ Comparative Studies of the
    different techniques
    ❏ Proposed Methodology
    ❏ Conclusions / Further Works
    ❏ References
    INTRODUCTION
    INTRODUCTION

    The emergence of web 2.0 is changing the world of social media. Not only online
    social media used to connect, share information and their personal opinion to others,
    but even business can also communicate, understand and improve their product and
    services through connecting in social media. The number of social media users
    increases every day and it is estimated in 2019 there will be up to 2.77 billion social
    media users worldwide.
    INTRODUCTION

    Social media is rich with raw and


    unprocessed data and the improvement in
    technology, especially in machine learning
    and artificial intelligence, allow the data to be
    processed and converted it into a useful data
    that they can benefit most business
    organization, especially in the domain of
    sentiment analysis
    INTRODUCTION

    Sentiment analysis is an
    approach that uses Natural
    Language Processing (NLP)
    to extract, convert and
    interpret opinion from a text
    and classify them into
    positive, negative or natural
    sentiment
    PROBLEM STATEMENT
    PROBLEM STATEMENT

    The main issues with sentiment analysis include the slowness of the models which
    in part may be due to the complexity of the text data, this includes the context of
    the words/phrases or sentences, the cultural background of the user, models trained
    on poorly prepared data, negation, sarcasm and irony, difficulty in determining the
    user’s stance. This can have an effect in industries like the Healthcare industry as
    this can endanger effective communication between patients and doctors since they
    will not understand patients’ needs and hence get to what services dissatisfied
    them, and may later on be unable to monitor the side-effects of medications based
    on social media posts and prevent unforeseen consequences, for example.
    OBJECTIVES

    OBJECTIVES
    OBJECTIVES

    MAIN OBJECTIVE

    ❏ The main objective of this paper is to obtain a fast and efficient model of sentiment
    analysis for social media.

    SPECIFIC OBJECTIVES

    ❏ To assess the existing techniques and compare their efficiencies.


    ❏ Investigate on what gave the following results and improve on existing
    deficiencies, bugs or hindrances discovered during this step
    ❏ After solving the issues in the previous objective, reassess the existing techniques,
    and later on based on the results, propose another technique or a blend of existing
    techniques in order to attain the main objective.
    LITERATURE REVIEW
    LITERATURE REVIEW

    There are basically two main approaches in terms of sentiment


    analysis which are :
    ❏ Lexicon-based approach
    ❏ Machine-Learning based approach
    Let us discuss what was done in each of those approaches in
    previous works
    LITERATURE REVIEW

    The work of Oscar Araque et al. (DepecheMood++: a Bilingual


    Emotion Lexicon Built Through Simple Yet Powerful Techniques
    (2019)) is an example of lexicon-based approach to sentiment
    analysis.
    LITERATURE REVIEW

    Here the main problem this work tackles is the low performance of
    lexicon-based approaches in terms of precision and coverage. In an
    attempt to solve this issue, this paper investigates whether
    computationally cheap techniques like document filtering, text pre-
    processing, & frequency cut-off can be used to improve the
    performance of rule-based techniques, as well as if machine
    learning techniques relying only on lexicon emotion scores can be
    used as the baseline for robust, complex and fast models which can
    be portable across languages.
    LITERATURE REVIEW

    To this end, a new lexicon was built upon the publicly available DepecheMood
    lexicon, which is generated from news sources distantly annotated with emotional
    scores. It was evaluated and released to the community as an extension of the
    original lexicon, built using a larger dataset, with a novel emotion lexicon targeting
    the Italian language and built with the same methodology. Experiments were then
    performed on six datasets/tasks exhibiting a wide diversity in terms of domain
    (namely: news, blog posts, mental health forum posts, Twitter), in different
    languages (English and Italian) and with different settings (both supervised and
    unsupervised), and task (regression and classification). The lexicon is called
    DepecheMood ++.
    LITERATURE REVIEW

    It is an upgrade on DepecheMood (which was built upon a dataset of 25.3k documents and
    13.5M words (530 words per document on average) built using an expanded dataset in
    order (i) to re-build the English lexicon on a larger corpus, and (ii) to build a novel lexicon
    targeting the Italian language.
    LITERATURE REVIEW

    First, a document-by-emotion matrix (MDE) per language is produced,


    containing the voting percentages for each document in the eight affective
    dimensions available in rappler.com for English(Happy, Sad, Angry,
    Annoyed, Don’t Care, Inspired, Afraid, Amused ) and the six available in
    corriere.it for Italian (Happy, Sad, Amused, Afraid, Angry / Annoyed,
    Inspired / Don’t Care).Then, the word-by-document matrices using
    normalized frequencies (MWD) are computed. That is, the number of
    occurrences of the word in a given document divided by the total number of
    words in that document. Previous experiments have shown this to be the
    best normalization option. After that, matrix multiplication are applied
    between the document-by-emotion and word-by-document matrices (MDE
    · MWD) to obtain a (raw) wordby-emotion matrix MW E. This method
    allows us to ‘merge’ words with emotions by summing the products of the
    weight of a word with the weight of the emotions in each document.
    LITERATURE REVIEW

    Several configurations were made to generate the DepecheMood+


    + lexicon. To fairly assess the performance of each configuration,
    randomly selected validation sets were employed, consisting of
    25% of the articles in our data for both rappler.com and corriere.it.
    For all the following evaluation experiments, the titles of such left-
    out sets were used. On a given headline, a single value for each
    affective dimension is computed by simply averaging the
    DepecheMood++ affective scores of all the words contained in the
    headline, this average value is then compared to the annotation for
    the headline.
    LITERATURE REVIEW

    The first configuration is Untagged Document Filtering where the


    performance of the lexicon on all documents(tagged and untagged)
    is compared to the performance on documents with a non-zero
    emotion annotation vector. It is evident that training with emotion-
    annotated documents and removing non-annotated ones leads to an
    improvement of results, which indicates that untagged documents
    add noise to the lexicon generation process.
    LITERATURE REVIEW

    The second configuration is Frequency Cutoff. Here different word


    frequency cutoff values are also explored to find a threshold that
    would remove noisy items without eliminating informative ones.
    LITERATURE REVIEW

    The third configuration is Term Weighting. Here the text


    representations are rather simple: an average of the affective scores
    of the words in the given text. With the aim of improving this, we
    propose an importance term weighting that makes use of the TF-
    IDF values to multiply the affect scores. TF-IDF (term frequency-
    inverse document frequency) is a statistical measure that evaluates
    how relevant a word is to a document in a collection of documents.
    This is done by multiplying two metrics: how many times a word
    appears in a document, and the inverse document frequency of the
    word across a set of documents.
    LITERATURE REVIEW

    Advantages of such a model include :


    ❏ Expanded training datasets which helped improve the
    performance
    ❏ Higher precision and coverage
    ❏ Based on computationally cheap techniques
    Weaknesses of such a model :
    ❏ Use of embeddings can improve the model (the technique
    map words that do not initially appear in a specific lexicon to
    a word that is contained in the lexicon mentioned above).
    LITERATURE REVIEW

    Machine-learning based approaches are discussed in the following


    paper: “ Sentiment Analysis of Social Networking Sites (SNS) Data
    using Machine Learning Approach for the Measurement of
    Depression ” by Hassan, Hussain, Husain, Sadiq, Lee.(2017).
    LITERATURE REVIEW

    The main problem tackled here is Depression, which has become


    the world’s fourth major disease. Compared with the high
    incidence, however, the rate of depression medical treatment is
    very low because of the difficulty of diagnosis of mental problems.
    According to the World Health Organization (WHO) survey in
    2012 more than 350 million people were suffering from depression
    and almost 1 million people with depression end their lives each
    year.
    LITERATURE REVIEW

    In that paper, different machine learning classifiers for sentiment classification were used
    since no single classifier is good for all kinds of datasets. There, they combined three
    different classifiers by using voting approach, in which each feature is assigned a number
    of votes and choose that label which gets the most votes.
    LITERATURE REVIEW

    In the proposed methodology, there were four main components that are
    preprocessing, feature extraction, meta learning and training data
    LITERATURE REVIEW
    LITERATURE REVIEW

    The models used here are Support Vector Machines, Naive Bayes,
    and Maximum Entropy.
    LITERATURE REVIEW

    The advantages of such an approach include :


    ❏ The use of multiple models, as there is no one-size-fits-all
    situation, helps in reducing errors, and makes the models
    faster and more efficient
    ❏ Since there are multiple models, the chance of accuracy is
    higher when talking about label classification
    LITERATURE REVIEW

    PROS OF Lexicon Based Approach

    ❏ Lexicon-based sentiment analysis methods are easily accessible as many publicly available resources (e.g.,
    SentiWordNet, DepecheMood) exist.
    ❏ They are less expensive because they do not require implementing advanced sentiment analysis algorithms.
    ❏ There is no need for training data, especially if companies use a dictionary-based approach, as the tags are
    determined manually, and there is quick access to the meaning of the words.

    CONS

    ● Lexicon-based sentiment analysis methods usually do not identify sarcasm, negation, grammar mistakes,
    misspellings, or irony. Thus, it may not be suitable for analyzing data gathered from social media platforms.
    LITERATURE REVIEW

    PROS

    ● Can be trained to detect sarcasm, irony, or negation in sentiment analysis. This can ease social
    media sentiment analysis.
    ● Learn the affective valence of the words, so they do not require a pre-determined dataset.
    ● Are faster than traditional sentiment analysis methods.
    ● Provide more accurate results.

    CONS

    ● Companies need a large or high-quality small dataset to have accurate classifications


    ● Noise (e.g., emojis, slang, or punctuation marks) can reduce accuracy
    ● Costs are higher compared to traditional, rule-based methods.
    PROPOSED METHODOLOGY
    PROPOSED METHODOLOGY
    ● First, data is collected from multiple sources e.g Twitter
    ● Then data is pre-processed (document filtering, term
    weighing, removal of stopwords etc…)
    ● Features of thé text are then extracted
    ● Data is then passed to text-based models like DepecheMood
    or VADER for firstus classifications
    ● Then data is passed to Machine Learning models such as
    Support Vector Machines, Naïve Bayes, Logistic Regression

    You might also like