Document-level Sentiment Analysis Based on Domain-specific Sentiment Words

Abstract

Aiming at the problem that the relationship between non-contiguous words in a sentence cannot be effectively captured, and the existing sentiment lexicon has poor adaptability in the field, a model for document-level sentiment analysis based on domain-specific sentiment words is constructed. We reconstructed word vectors using attention mechanism to capture the relationship between non-contiguous words in word vectors; Words are synthesized using Asymmetric Convolutional Neural Network. Sentences are synthesized by Bidirectional Gated Recurrent Neural Network based on attention mechanism to form document vector features; we used CNN to construct a domain-specific sentiment dictionary to generate emotional vector features; Document vector features and emotional vector features are combined using a linear binding layer to form document features that facilitate document classification. By comparing the performance of this method with other methods through experiments, the results show that there is a big advantage in classification accuracy, and can be widely used in various specific fields such as public health.