Context-Sensitive Neural Sentiment Classification

Enabling a computer to precisely understand a document so that it can predict the sentiment is crucial, yet the unsolved goal of NLP! Although, the majority of existing methods classify text documents only based on semantic of text data and ignore relevant contextual information. There are a few studies which take this data into account, but they suffer from model complexity. To address this issue, we propose an attention-based hierarchical neural network model to incorporate user preferences and product characteristics into sentiment classification task. Our model benefits from both character and word level embedding to build a hierarchical representation of the documents content. Afterward, contextual information via attentions mechanism applied over word encoding layer to capture the user-specific meaning of each word. We evaluate our system on two publicly available dataset Yelp and Amazon for fine-grained sentiment classification. Our model shows significant performances over several robust baseline methods.