Advanced Rule-Based Approaches for Explainable Sentiment Analysis (for Czech)

This thesis deals with implementing a rule-based shifter pattern approach for sentiment analysis for the Czech language and comparing it to other modern approaches, mainly the state-of-the-art deep neural network approach. Two algorithms are implemented based on existing research for other languages, one using a proximity approach and the other using dependency trees. Lists of Czech shifters for different patterns like intensificators and deintensificators are collected. Existing sentiment lexicons are used alongside a method for automatic lexicon generation. For evaluation, apart from already existing Czech datasets, two new datasets from an economics domain are created, utilizing automation with generative artificial intelligence for parts of the process. The evaluation shows that the rule-based approach didn't manage to outperform others, instead falling behind. The difference between the best rule-based setup and the best state-of-the-art DNN approach in terms of macro F-score was between 0.14 and 0.23. Some of the main causes of errors included the imperfections of existing Czech lexicons, ambiguity, and complex semantics that are hard to capture with a limited set of rules or not capturing a certain pattern due to an incomplete list of possible words. The proximity approach performed slightly better than the dependency tree one, which still has room to improve. Compared to neural-network approaches, the rule-based approach showed itself as more explainable, faster in execution speed, and less resource-dependent, however, it would require substantial effort to bring it close to the level of the current state-of-the-art in terms of accuracy.