Institutional Repository of Vilnius University
Dezinformacijos identifikavimas tekstuose /
Abstract
dc:descriptionDisinformation is a constantly growing problem that is becoming impossible to solve manually. Therefore, models that are able to solve the problems of natural language processing are applied in this work to identify misinformation. In this paper, the transformer architecture and the GPT-2 and BERT architectures based on it were analyzed. Two scientific papers were also analyzed which attempted to develop similar models that could classify true and false texts as accurately as possible. During the work, two Lithuanian data sets were collected, which were used to train the GPT-2 model to generate Lithuanian text. A total of 4 models generating Lithuanian texts were created, where one of the models achieved 38 % accuracy. The text generated by these models was later used to train the disinformation classification model. This paper reviews the operation of these models, their metrics and the generated text. 3 classification models based on the multilingual BERT model were further developed. The paper analyzed the ability of these models to recognize the text generated by different models, as well as text written by people. The work shows that it is important to have a quality text generation model, because the training results of the classification model depend on it.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Sabalys, Benas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD146242328&prefLang=en_US