The influence of lexicon and punctuation on cross-lingual transfer performance of text-based emotion recognition language models

Automatic emotion recognition in text is a popular task in the field of natural language processing. One of the main problems researchers face is the lack of labeled emotional data in many languages – to address this issue, cross-lingual transfer of models pretrained on sufficient data from another language is often used. Most existing studies on this topic primarily focus on improving classification algorithms, while the influence of linguistic factors of the languages involved remains underexplored. The subject of this research is textual emotional data used in the task of automatic emotion recognition. The aim of this study is to examine the influence of punctuation and lexical similarity on the performance of cross-lingual model transfer for emotion recognition. As a method, an experiment was conducted in which the XLM-RoBERTa model was trained on the BRIGHTER dataset, specifically on data from German, Spanish, Chinese, Hindi, and Tatar languages, and the trained model was subsequently tested on Russian language data. This work presents a detailed experimental analysis of cross-lingual transfer of textual language models for automatic emotion recognition and establishes the influence of lexicon and punctuation on the quality of their transfer. The results showed that the similarity between the lexicon of the source and target languages — primarily semantic similarity — is a key factor for achieving good emotion classification performance in cross-lingual transfer, whereas punctuation and emojis play a secondary role. Nevertheless, punctuation may still significantly influence transfer performance when it's predominantly associated with a single emotion. The results obtained in this study may be used in the creation of new emotional text datasets as well as in improving existing ones.