Aspect-based Sentiment Analysis using Hierarchical Attention Networks

Abstract

The proliferation of user contributed content in e-commerce, social media, and review sites have increased the demand for more accurate approaches to polarity detection. This need is met by Aspect-Based Sentiment Analysis (ABSA) that provides sentiments related to certain aspects of an entity which is useful for e-commerce, healthcare, and social media analysis. In this work, we propose the use of Hierarchical Attention Networks (HAN) for ABSA due to their ability to model at multiple granularities to enhance aspect-level sentiment classification. The presented framework improves position-sensitive embeddings and uses multi-head attentions to boost the scalability, interpretability, and performance of the cased model on multilingual and domain-shifted datasets. On various datasets, the proposed method showed better accuracy (93%), precision (92%), and F1-score (92%) than conventional methods. The model also performs well in dealing with implicit features and subtle sentiment patterns, which is further accompanied by attention visualization for better understanding. This research establishes new standards in ABSA, solving scalability and domain adaptability issues and opening the way to its application in large-scale sentiment analysis.