Aspect based sentiment analysis in Turkish

Target-based sentiment analysis enables identifying positive, negative, or neutral sentiment for each aspect term in a sentence, allowing the analysis of multiple opinions. In multi-aspect scenarios where traditional sentiment analysis falls short, ABSA provides a more precise understanding of user feedback. In this context, research tailored to the Turkish language is crucial to bridge the gap in existing resources and enable more effective sentiment-aware applications. However, research in this area for the Turkish language is limited and needs further development. This thesis introduces a new open-domain dataset of 6,000 Turkish texts from domains such as product reviews, accommodation, education, and social media, annotated with single- and multi-word aspect terms. The ABSA task is addressed through two subtasks: aspect term extraction and aspect-level sentiment classification. Aspect extraction is modeled as a Named Entity Recognition (NER) problem using the BIOS tagging scheme. A learnable fusion model combining contextual embeddings from BERT and ELECTRA achieves 88.53% accuracy and 88.52 F1-score. Sentiment classification reaches 86.97% accuracy and 83.69 F1-score. This study provides a valuable open-domain resource including aspect terms, sentiment polarities, and category information, contributing to Turkish ABSA research.