Exploring Different Models for Aspect-Based Sentiment Analysis

Understanding what people think about specific parts of products or services is crucial, especially in the online business world. Early attempts at sentiment analysis lumped opinions into overall document or sentence categories, missing the finer points. Aspect-based sentiment analysis (ABSA) stepped in to link sentiments to particular aspects mentioned in reviews. ABSA is relatively new and faces three big hurdles: it's often tailored to specific fields, needs labeled data, and hasn't fully explored the potential of newer large language models (LLMs) like GPT, PaLM, and T5. To tackle these, we looked at various datasets including DOTSA, MAMS, and SemEval16, and tested models like ATAE-LSTM, flan-t5-large-absa, DeBERTa, PaLM, and GPT-3. 5-Turbo. DeBERTa showed strong performance across the board, while PaLM was highly competitive for aspect term sentiment analysis (ATSA) tasks. PaLM also did well in all domains tested—restaurants, hotels, books, clothing, and laptops. Our study showed that these models' performance varies by domain, which is important to know for both ATSA and aspect category sentiment analysis (ACSA) tasks. These findings help us grasp where these models excel and where they fall short, guiding future ABSA research and development.