Interactive POS-aware network for aspect-level sentiment classification

Abstract

Existing aspect-level sentiment-classification models completely rely on the learning from given datasets. However, these are easily misled by biased samples, resulting in learning some ill-suited rules that limit their potential. The information of some specific part-of-speech (POS) categories often indicates the word sentiment polarity, which can be introduced as prior knowledge to facilitate prediction of the model. Accordingly, we propose an interactive POS-aware network (IPAN) that explicitly introduces the POS information as reliable guidance to assist the model in accurately predicting sentiment polarity. We distinguish the information of different POS categories using a POS-filter gate and reinforce the features extracted from adjectives, adverbs, and verbs via a POS-highlighting attention mechanism. This enables the model to concentrate on the words that contain significant sentiment orientations and to obtain the most practical learning experience. To emphasize the target information, we construct a target-context gate that enables the interaction of the target information with contexts; consequently, the model considerably focuses on target-related sentiment features. The experiments on SemEval2014 and Twitter datasets verify that our IPAN consistently outperforms the current state-of-the-art methods.