Aspect-based sentiment analysis for turkish using deep learning model combinations

People use the World Wide Web heavily to share their experiences with entities such as products, services or travel destinations. These reviews and comments are essential for consumer decisions and create a valuable source that may be used to measure satisfaction. Sentiment analysis (SA) is a natural language processing task of identifying opinions expressed in such text fragments. Aspect-Based Sentiment Analysis (ABSA) is a research area whose ability to provide fine-grained insights into sentiment expressions related to specific aspects or features of entities and it has gained significant attention in recent years. In this thesis, we mainly focus on sentiment analysis and aspect extraction parts of ABSA especially in Turkish. For SA, we used two approaches. The first is to develop average review vectors from noncontextualized and contextualized embeddings and add weights to these review vectors using word frequencies in positive and negative sensitivity-tagged reviews. The second is to use different combinations of deep learning models with different embedding methods and tokenization approaches. We make a comprehensive analysis of eleven different tokenization approaches, including the commonly used subword methods and morphologically motivated segmentations. We ensemble the models to find out which tokenization approaches produce better results together. The final part of the thesis is related to ABSA's crucial aspect extraction component, which involves identifying and extracting aspect terms from reviews. We develop a novel tree positional embedding, in addition to embeddings for tokens and Part-of-Speech (POS) tags, and deep learning algorithms are used to develop a novel learning model. A novel dataset is created from English reviews by translating them into Turkish and relabelling them in addition to an existing Turkish review dataset.