Exploring Hierarchical Linguistic Structure for Aspect-Based Sentiment Analysis

Abstract

The primary aim of aspect-based sentiment analysis is to extract the sentiment of a specific aspect from an opinion textual document. Generally, text segments relevant to an aspect are first localized. Using lexicon-based approaches, polarity scores from terms in the segments are subsequently combined into an overall score. In this paper, we propose a new score aggregation strategy that utilizes linguistic structures in several ways. A given textual document is segmented into elementary discourse units (EDUs) with relations between them. Polarity scores for EDUs are then computed from all aspect-related terms, identified using term dependency structure. The EDU scores are hierarchically combined into the scores for their local aspect segments and subsequently into those for aspect segments. Experimental results on online product reviews demonstrate that our new score aggregation method and EDU-level score calculation, exploiting term dependencies, improve sentiment classification at the level of local aspect segments.