dependency-attention-based lstm for target-dependent sentiment analysis

Abstract

Target-dependent sentiment analysis is a fine-grained sentiment analysis and has received an increasing attention. For target-dependent sentiment analysis, the key issue is to capture the important context information according to the given target word. While some critical information in the context may be in a long distance from the target word, so it is significant to explore how to adequately and directly capture these long-range information. The dependency relation can connect words which are relevant in syntax but far in word order. Inspired by this, we propose Dependency-Attention-based Long Short-Term Memory Network (DAT-LSTM) and Segmented Dependency-Attention-based Long Short-Term Memory Network (Seg-DAT-LSTM) for target-dependent sentiment analysis. The dependency-attention mechanism utilizes dependency relation to fully capture long-range information for certain target. Experiments on the tweet dataset and SemEval 2014 dataset indicate that our models achieve state-of-the-art performance on target-dependent sentiment classification.