Opinion Units: Concise and Contextualized Representations for Aspect-Based Sentiment Analysis
Abstract: We introduce opinion units, a novel approach to Aspect-Based Sentiment Analysis (ABSA) that extends traditional aspect-sentiment pairs by including substantiating excerpts derived through hybrid abstractive-extractive summarisation. This reduces the information loss inherent in traditional ABSA methods, and the structured format facilitates downstream processing tasks. Experiments on benchmark datasets for ABSA demonstrate that large language models (LLMs) can accurately extract opinion units using a few-shot approach. The main types of errors are overlooking aspects in the text, and characterising objective statements as opinions. The method eliminates the need for labelled data and allows the LLM to dynamically define aspect types. Additionally, we present a case study on similarity search for opinions in academic datasets and public review data. Our results indicate that searches based on opinion units are more successful than those using traditional data-segmentation strategies, demonstrating robustness across datasets and embeddings.
Paper Type: Long
Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining
Research Area Keywords: argument mining;retrieval; zero/few-shot extraction;stance detection;extractive summarisation
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 2566
Loading