Evaluating Lexicon-Based Models versus BERT for Sentence-Level Sentiment Analysis in Swedish

Abstract

This thesis explores the development and evaluation of different approaches to sentiment analysis for the Swedish language, focusing on sentence-level sentiment detection. The study compares traditional rule- and lexicon-based models with modern machine learning approaches, particularly the Bidirectional Encoder Representations from Transformers (BERT), as well as a hybrid model combining the rule-based model with Support Vector Machines SVM. Utilizing the Sparv pipeline for linguistic analysis and breadkdown in tandem with the sentiment lexicon SenSALDO, we aim to enhance the existing research on Swedish rule-based models by inclusion of linguistic features. The research also involves expanding the lexicon with neutral, positive and negative entries in order to improve coverage and accuracy of sentence level sentiment analysis. The evaluation highlights the strengths and weaknesses of each model where the BERT model was the best performing overall, especially for neutral sentences, while the rule based and hybrid model were much better at positive sentences, for negative sentiment detection the hybrid SVM model was the best performing. Our thesis contributes to the ongoing discourse on effective sentiment analysis in non-English languages and offers insights for further advancements in natural language processing (NLP) for Swedish.