Use of transformer models in the analysis of sentiments of textual sequences in the Croatian language

Zbog trenutne aktualnosti teme COVID-19 pandemije u ovom radu koristi se predtrenirani trojezični CrosloEngual model za analizu promjene sentiment na hrvatskom podatkovnom podskup prikupljenom prije početka COVID-19 pandemije i te podskupu kreiran neposredno nakon početka pandemije. Pošto navedeni skup podataka je neoznačen prvotno se model fino podešava na sličnom skupu slovenskih i hrvatskih novinskih članaka te se zatim klasificira i analizira sentiment prikupljenog COVID-19 skupa podataka. Završna analiza upućuje na moguća ne podudaranje skupova podataka za treniranje i COVID-19 skupa podataka zbog velikih ne uravnoteženosti između distribucija klasa predviđenog sentimenta na COVID-19 skupu podataka.Due to the current popularity of the COVID-19 pandemic, this paper uses a pre-trained trilingual CrosloEngual model to analyze sentiment change in the Croatian data subset collected before the COVID-19 pandemic and the subset created immediately after the pandemic. Since the mentioned dataset is unmarked, the model is initially fine-tuned on a similar set of Slovenian and Croatian news articles, and then the sentiment of the collected COVID-19 data set is classified and analyzed. The final analysis suggests a possible mismatch of the training datasets and the COVID-19 dataset due to large imbalances between class distributions on predicted sentiment of the COVID-19 dataset