Learning for target-dependent sentiment based on local context-aware embedding

Target-dependent sentiment classification is a fine-grained task of natural language processing to analyze the sentiment polarity of the targets. In order to address the difficulty of locating important sentiment information of targeted sentiment classification, recent research mostly applies attention mechanisms to capture the information of important context words, while the attention mechanism is subject to many drawbacks, e.g., dependent on network architecture and expensive. Recent studies show the significant effect of the local context focus (LCF) mechanism in capturing the relatedness between a target’s sentiment and its local context. However, the LCF simply applies the fusion of global and local context features to classify sentiment, neglecting to empower the network to be aware of deep information of local context. In this paper, we propose a novel local context-aware network (LCA-Net) based on the local context embedding (LCE). Moreover, accompanied by the sentiment classification loss, the local context prediction (LCP) loss is proposed to enhance the LCE. The experimental results on three commonly used datasets, i.e., the Laptop and Restaurant datasets from SemEval-2014 and a Twitter social dataset, show that all the LCA-Net variants achieve promising performance improvement compared to existing approaches in extracting local context features. Besides, we implement the LCA-Net with different neural networks, validating the transferability of LCA architecture.