Naslov
How good BERT based models are in sentiment analysis of Croatian tweets: comparison of four multilingual BERTs

Autori
Ptiček, Martina

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), znanstveni

Izvornik
Proceedings of 32nd International Scientific Conference Central European Conference on Information and Intelligent Systems / Vrček, Neven ; Pergler, Elisabeth ; Grđ, Petra - Varaždin : Fakultet organizacije i informatike Sveučilišta u Zagrebu, 2021, 175-182

Skup
32nd Central European Conference on Information and Intelligent Systems (CECIIS 2021)

Mjesto i datum
Varaždin, Hrvatska, 13.10.2021. - 15.10.2021

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
Sentiment analysis, contextual word embeddings, multilingual BERT, Croatian language

Sažetak
Contextual word embeddings like BERT or GPT give the state-of-the-art results in a vast array of tasks in NLP - especially when applied to English datasets, given the fact that these models themselves were trained on numerous data in English language. However, the successfulness of these models has not yet been sufficiently researched for low resource languages, as Croatian. This paper describes a comparison between the application of BERT based multilingual word embeddings (mBERT, DistilBERT, XLM-RoBERTa, CroSloEngual) in sentiment analysis on tweets in Croatian language. The article shows that BERT based multilingual models give good results in sentiment analysis in Croatian language, particularly the models trained on larger sets of data in Croatian as XLM-RoBERTa and CroSloEngual.

Izvorni jezik
Engleski

Znanstvena područja
Informacijske i komunikacijske znanosti

POVEZANOST RADA