Exploring linguistic structure for aspect-based sentiment analysis

Aspect-based sentiment analysis is a task that extracts relevant sentiments of a specific aspect. An opinion text is usually composed of views on different aspects of an entity. By investigating the sources of errors, we observe that a scoring method at the level of elementary discourse units (EDUs) highly contributes to the accuracy of sentiment classification at the aspect level. Score aggregation can be improved by considering linguistic structures between EDUs in a hierarchical manner. We propose a new score aggregation strategy that incrementally aggregates sentiment scores from EDUs to local segments and from local segments to an aspect. The experimental results on a product review dataset demonstrate that our new score aggregation method improves the performance of sentiment classification at the aspect level. At the EDU level, calculation of polarity scores using an all-term average yields better performance compared to score calculation based on opinion phrases extracted by using term dependencies.