Comparative analysis of methods for creating a sentiment lexicon of the Serbian WordNet

What is it about?

Automatic sentiment analysis represents a key segment in the domain of natural language processing, with wide applications in various fields such as social networks, marketing, and public opinion monitoring. SentiWordNet, a lexical resource created for sentiment analysis, assigns numerical sentiment values to synsets from WordNet. Although initially developed for English, existing adaptations have enabled its global application in sentiment analysis in various languages. Within the BalkaNet project, a Serbian WordNet was developed. The values from SentiWordNet were transferred to the Serbian WordNet through direct mapping to corresponding synsets from Princeton's WordNet, but the results obtained with this approach were not satisfactory. In this paper, we analyzed how SentiWordNet was initially constructed. The sentiment analysis of synsets was reduced to gloss classification. The gloss was vectorized using the bag-of-words approach. Further classification was performed using a combination of support vector machine and naive Bayes methods. This paper examined how these methods can be improved and adapted for the Serbian WordNet, to obtain a lexicon whose sentiment values better reflect the real sentiment in the Serbian language. The same methods were used, support vector and naive Bayes, with minor modifications, replacing support vectors with ADA boost, as a complete replacement with more modern methods: recurrent neural networks, transformers, and large language models finely tuned for sequence classification. Thus, multiple lexicons like SentiWordNet were created. The paper also describes software tools developed for this purpose.

Featured Image