Building an emotion lexicon for Serbian using curated language resources

Abstract

This article introduces a methodology for developing the first emotional affect lexicon for the Serbian language. The proposed methodology involves leveraging a Large Language Model (LLM), specifically the GPT-3-based gpt-3.5-turbo and GPT-4-based gpt-4.1 models, in conjunction with the Serbian WordNet language resource to align the English lexicon with Serbian-specific morphological and linguistic characteristics. The effectiveness of the Serbian emotion lexicon (EmoLex.SR), comprising 13,584 affective words, has been validated through emotion detection experiments using emotion-annotated corpora. The experiments demonstrated outstanding performance compared to the NRC lexicon automatically translated into Serbian, achieving a macro F1 score of 74.4% for sentences written in Serbian. In particular, the lexicon outperforms its automatically translated counterpart in detecting emotional categories across three distinct datasets, with an average improvement by 14.7% in terms of macro F1 score. The development of the EmoLex.SR lexicon and the accompanying annotated parallel corpora, referred to as LLM-Emo.SR, extends the emotion detection capabilities for Serbian language processing. This enables a more accurate interpretation of emotions in Serbian text and enhances Natural Language Processing applications for the Serbian language. Although the methodology for creating the lexicon is demonstrated for Serbian, it can also be successfully applied to other languages. The lexicon is made publicly available to the scientific community for use and further refinement.