Effective-target Representation via LSTM with Attention for Aspect-level Sentiment Analysis
Quan Liu, Hiroaki Mukaidani · 2020
Target-dependent sentiment analysis aims to classify the sentiment polarities of a given target in its comment. Previous work has recognized the importance of the target in aspect-level sentiment analysis and focused on modeling relationships between targets with context. However, these studies always used averaging the hidden output of the target context as target representation. In some cases, this approach may not be appropriate, because different target words usually do not contribute equally. In this work, an attention mechanism to compute the importance of target words based on each left context word and right context word is proposed. As a result, effective- target representation can be used to capture the words most contributing sentiment in context. In particular, two gates to control the effects of left and right context information on the target polarities' tendency are added as a novel concept. To demonstrate the effectiveness of the proposed scheme, an experiment on restaurant datasets was performed.