Improving Multi-Category Sentiment Classification by Semantic Dual-Granularity and Syntax-Path Encoding

Abstract

Semantic features, word sequences, and syntactic structures are three key elements for human classification of multi-category sentiment texts. However, current sentiment classification neural network models focus mostly on semantic features and word sequences, thereby losing syntactic structure information in modeling and limiting classification accuracy. In this paper, we combine three key elements at different levels and propose an enhanced model with semantic dual-granularity and syntax-path encoding for multi-category sentiment classification. In particular, at the semantic level, we utilized two long short-term memories to process semantic dual-granularity units that are word and n-gram sequences; then we combined the two memory outputs via the attention mechanism to highlight keywords or key phrases in sentences. At the syntactic level, we extracted syntactic features by encoding each word's syntax path into a syntactic neural network, aiming to capture implied sentiments through an expression pattern. Finally, the two levels' models were combined for multi-category sentiment classification. We applied our model to three common and public multi-category sentiment datasets. Experimental results showed our model significantly outperforming eight strong baseline models with a maximum accuracy improvement of 48.8%; a minimum accuracy improvement of 1%; and average accuracy improvements of 20.9%, 19.3%, and 14.8% on each dataset, respectively.