Department of General Linguistics

Cikkünkben bemutatjuk a laikusoknak címzett hivatalos szövegek osztályozási kísérletét felügyelt gépi tanuló algoritmusok segítségével. Vizsgálatunkhoz szakértők által, kézzel készített korpuszt használunk, amely közérthetőre fogalmazott és még átfogalmazás előtt álló mondatokat tartalmazott. Célunk ezzel egy olyan gépi tanult modell készítése, amely alkalmas lehet a szakértők figyelmét felhívni azon mondatokra egy-egy hivatalos szövegben, amelyek további megfontolást érdemelnek a szöveg közérthetőbbre alakítása során, ezzel gyorsítva a szakértői munkát. A kísérletet pilot jelleggel végezzük, az eredmények függvényében korszerűbb módszerek (pl. LSTM, BERT) esetleges kipróbálása előtt, a tapasztalatokat pedig a fentiek szerinti bináris klasszifikációs problémára jellemzően értékeljük.

The current article briefl y presents a pilot machinelearning experiment on the classifi cation of offi cial texts addressed to lay readers with the use of support vector machine as a baseline and fastText models. For this purpose, a hand-crafted corpus was used, created by the experts of the National Tax and Customs Administration of Hungary under the offi ce's Public Accessibility Programme. The corpus contained sentences that were paraphrased or completely rewritten by the experts to make them more readable for lay people, as well their original counter pairs. The aim was to automatically distinguish between these two classes by using supervised machine-learning algorithms. If successful, such a machine-learning-based model could be used to draw the attention of experts involved in making the texts of offi cial bodies more comprehensible to the average reader to the potentially problematic points of a text. Therefore, the process of rephrasing such texts could be sped up drastically. Such a rephrasing (considering, above all, the needs of the average reader) can improve the overall comprehensibility of offi cial (mostly legal) texts, and therefore supports access to justice, the transparency of governmental organizations and, most importantly, improves the rule of law in a given country.

The current article briefly presents a pilot machine-learning experiment on the classification of official texts addressed to lay readers with the use of support vector machine as a baseline and fastText models. For this purpose, a hand-crafted corpus was used, created by the experts of the National Tax and Customs Administration of Hungary under the office’s Public Accessibility Programme. The corpus contained sentences that were paraphrased or completely rewritten by the experts to make them more readable for lay people, as well their original counter pairs. The aim was to automatically distinguish between these two classes by using supervised machine-learning algorithms. If successful, such a machine-learning-based model could be used to draw the attention of experts involved in making the texts of official bodies more comprehensible to the average reader to the potentially problematic points of a text. Therefore, the process of rephrasing such texts could be sped up drastically....

In this paper, a brief study will be presented with regard to the issue of Named Entity Recognition (NER) in legal texts. To get an overall picture, we examined closely the output of two existing analysers: the "magyarlanc" linguistic processing toolkit [1] and a Named Entity Recognition system developed by the Natural Language Processing Group at the University of Szeged [2]. Firstly, short references are made to named entity recognition projects in the literature considered important in the current framework. Secondly, quantitative analyses of the data will be presented. At the end of the study, some problematic cases and potential solutions will be discussed which will be followed by the discussion of the future research.

This paper outlines the ParlaMint project from the perspective of its goals, tasks, participants, results and applications potential. The project produced language corpora from the sessions of the national parliaments of 17 countries, almost half a billion words in total. The corpora are split into COVID-related subcorpora (from November 2019) and reference corpora (to October 2019). The corpora are uniformly encoded according to the ParlaMint schema with the same Universal Dependencies linguistic annotations. Samples of the corpora and conversion scripts are available from the project's GitHub repository. The complete corpora are openly available via the CLARIN.SI repository 1 for download, and through the NoSketch Engine 2 and KonText 3 concordancers as well as through the Parlameter 4 interface for exploration and analysis.

A dolgozatban bizonyos pragmatikai és szemantikai sajátságokat vizsgálunk magyar nyelvű, nagy méretű spontánbeszéd-korpusz (StaffTalk) alapján. A vizsgálati korpusz is hétköznapi szituációkban, külső hatásoknak is kitett munkahelyi környezetben, spontán módon létrejött nyelvi tartalmakból áll, vagyis a kutatásban résztvevők szabadon megválaszthatták beszélgetésük tárgyát, hosszát és partnereit. A korpusz létrehozása során a hanganyagokat legépelték, majd azt követően számos szempont alapján annotálták. A jelen vizsgálatokat ezeket az annotációkat felhasználva végezzük el.

The growing number of digitally accessible text corpora and the accelerating development of NLP tools and methods (particularly the emergence of powerful large-scale language models) have allowed their widespread use in various classification tasks, including the vast field of sentiment analysis. However, these models must often be fine-tuned to perform this task efficiently. Therefore, we aimed to create a transformer-based fine-tuned model for the emotion and sentiment analysis of Hungarian political texts. The training data for the model were the manually annotated parliamentary speech texts from 2014 to 2018, which have the advantage of being rich in various emotions. The compiled corpus can be freely used for research purposes. In our work, we describe in detail the process of fine-tuning the Hungarian BERT model for sentiment and emotion classification, the performance achieved, and the typical classification errors, mainly due to a lack of recognition of pragmatic and other language use features by the fine-tuned models. INDEX TERMS Fine-tuned BERT-model, huBERT, emotion analysis, sentiment analysis, political communication.

Harnessing the power of Deep Learning is becoming commonplace nowadays, and the legal field is no exception. Most solutions embrace supervised approaches that require a lot of labeled data. Active Learning is a technique that exploits the already labeled data hence optimizing the time spent for labeling, saving time and costs. In this paper, the efficiency of Active Learning is investigated on a real-world hierarchical, multi-class multi-labeled dataset consisting of Hungarian legal documents from the taxation domain. Three different vectorization techniques (TF-IDF, fastText, BERT) are compared. The results suggest that by applying Active Learning, 40-60% less data is sufficient compared to random sampling and that lowering the batch size in an active learning cycle has a modest effect on reaching optimal performance.

One crucial aspect of access to justice and access to legal information is the comprehensibility of legal text. The complexity and specialized terminology of legal language often prevents citizens from understanding legal texts and representing themselves effectively in legal proceedings. The new developments of Machine Learning, such as solutions based on Large Language Models, could represent a significant advancement in access to legal information, as they can transform complex legal texts into more straightforward, more understandable forms for laypeople. This paper attempts to exploit the capabilities of OpenAI’s GPT-4 model to produce automatic Plain Language transcriptions of legal texts. The experiment concerns four specific linguistic features, and the results are analyzed manually from both a legal and a linguistic point of view.

Automated sentiment analysis of textual data is one of the central and most challenging tasks in political communication studies. However, the toolkits available are primarily for English texts and require contextual adaptation to produce valid results—especially concerning morphologically rich languages such as Hungarian. This study introduces (1) a new sentiment and emotion annotation framework that uses inductive approaches to identify emotions in the corpus and aggregate these emotions into positive, negative, and mixed sentiment categories, (2) a manually annotated sentiment data set with 5700 political news sentences, (3) a new Hungarian sentiment dictionary for political text analysis created via word embeddings, whose performance was compared with other available sentiment dictionaries. (4) Because of the limitations of sentiment analysis using dictionaries we have also applied various machine learning algorithms to analyze our dataset, (5) Last but not least to move towards state-of-the-art approaches, we have fine-tuned the Hungarian BERT-base model for sentiment analysis. Meanwhile, we have also tested how different pre-processing steps could affect the performance of machine-learning algorithms in the case of Hungarian texts.