Research On Aspect-level Sentiment Classification Based On Dependency Tree And Attention Network

Aspect-level sentiment classification is a fine-grained task in the field of sentiment analysis,which aims to determine the sentiment polarity of the sentence towards the aspect after given a sentence and an aspect.Due to the sentiment polarity of an aspect needs to consider both the contexts information and the aspect information,the key element in the aspect-level sentiment classification task is to characterize the relationship between the aspect and the contexts.In order to solve the problem,the attention mechanism regards the aspect as the attention calculation objective,so that the attention model can learn the association between aspect and contexts directly,thereby different sentiment representations are generated for each aspect.The attention mechanism has achieved excellent performance and has become the state-of-the-art methods in the aspect-level sentiment classification task.Currently,the aspect representations simply use the word embedding of the aspect with the attention mechanism,which leads the attention model unable to learn the appropriate attention weights.This paper proposes a dependency subtree attention networks(DSAN)model to solve the problem.The DSAN model firstly extracts the dependency subtree(aspect sub-sentence)that contains the modification information of the aspect based on the dependency tree of the sentence,and then utilizes a bidirectional GRU network to generate an accurate aspect representations,and uses the dot-product attention function for the aspect represented by the dependency subtree manner,thereby the appropriate attention weights can be learned.In addition,the DSAN model also regards the aspect and context words syntax distances in the dependency tree of the sentence as the input of the aspect position information,and provides an aspect-customized contextual representations for each aspect that is weighted by the position,which can help the model to predict respective sentiments of multiple aspects in a common sentence more flexibly.This paper also trains domain-specific word embeddings to make further improvement in the performance of the DSAN model for aspect-level sentiment classification.The experimental results showed that the DSAN model achieved the comparable state-ofthe-art performance in the two benchmark datasets of the SemEval 2014.Moreover,the DSAN model achieved state-of-the-art performance after using Word2vec-skipgram word embeddings that was trained by domain-specific corpus,gained 76.96% accuracy of the Laptop dataset and 81.43% accuracy of the Restaurant dataset.