Expressing emotions is an important part of human communication, which is why development of automatic emotion detection has accelerated in recent years. Emotion information is useful in advertising, psychology, human–robot interaction, and related areas. No model for emotion detection from Slovene text was available, so we developed one using modern natural language processing methods. As no annotated Slovene training set exists, we relied on cross-lingual transfer, training the trilingual CroSloEngual BERT model on an English corpus. For evaluation, we created the SloReddit dataset ($162$ comments). With zero-shot cross-lingual transfer we obtained the macro $F_1$ score of $0.657$ on this set. Fine-tuning the same model on roughly one hundred examples from it raises the score to $0.674$ using cross-validation, which is comparable to similar emotion recognition models for English text.
We show that roughly one hundred manually labelled native examples contribute more than fine-tuning on thousands of translated ones.
We also compared monolingual SloBERTa, an ensemble of binary classifiers, and prompting of GaMS models. None of these additional approaches surpassed fine-tuned CSE-BERT on native comments.
|