Institutional Repository of Vilnius University
mBERT modelio taikymas tekstų klasifikavimui pagal sentimentą lietuvių kalboje /
Abstract
dc:descriptionThe dramatic increase in the volume of user-generated content on the internet makes it particularly important for various public and business actors to analyse consumer sentiment. The main goal of this thesis is to test the capability of the mBERT model to perform the task of text classification based on sentiment in the Lithuanian language, i.e. to classify a text according to the predominant emotion it contains as positive or negative. Three different English datasets were chosen for this work, each of which was translated into Lithuanian using Google and DeepL translators. Each dataset was translated using two translators, so there are 3 different versions of each dataset - the original (English), and two Lithuanian versions. This way the classification performance of the mBERT model is compared between the English and the Lithuanian text. In this work, 9 experiments were conducted to fine-tune and test the mBERT model. The performance of the model was measured using metrics such as accuracy, recall, F_1 value, precision and AUC-ROC. The experimental results showed the highest accuracy of 87% of the model in classifying the Lithuanian text based on its sentiment into positive and negative. This is a significantly higher classification accuracy compared to traditional NLP techniques but is 3% lower than using English language datasets. All model performance parameters were similar to accuracy. The choice of translator did not have a significant impact on the model performance. This work demonstrates the significant potential of the mBERT model in Lithuanian, but also highlights further potential improvements to the dataset and the model fine-tuning process.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kučinskas, Karolis,
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:ELABAETD210644526&prefLang=en_US