Filter gate network based on multi-head attention for aspect-level sentiment classification

Abstract

Aspect-level sentiment classification (ASC) is an important branch of sentiment analysis. Its purpose is to identify the sentiment polarity of the context for a given aspect. In recent years, ASC has received widespread attention from more and more researchers. Most of the previous work used word embedding, which was trained from large corpora, as context representation. However, this method cannot fully express the actual meaning of the context. In addition, the attention mechanism in ASC brings noise and captures context words that are irrelevant to the current aspect. Based on the above problems, we propose a novel neural network, named Filter Gate Network based on Multi-head attention (FGNMH). First, we train the context in a domain-specific corpus and integrate the part-of-speech features of the context to enrich the representation of the context. Second, we use multi-head attention mechanism to model contextual semantic information. Finally, a filter layer is designed to remove context words that are irrelevant to current aspect. To verify the effectiveness of FGNMH, we conduct a large number of experiments on SemEval2014, Restaurant15, Restaurant16 and Twitter. Experimental results show that FGNMH can reach 83.92%, 81.24%, 84.05%, 85.37%and 74.37% on five data sets, respectively.