Interactive Attention Encoder Network with Local Context Features for Aspect-Level Sentiment Analysis

Abstract

The sentiment analysis of the data collected by Internet of Things devices has been widely concerned by researchers. Aspect sentiment analysis is a fine-grained task of sentiment analysis, which aims to identify the sentiment polarity of specific aspects in a given context. Most of the previous studies have modelled context and aspect terms using Recurrent Neural Network (RNN) and attention mechanisms. However, the RNN model is difficult to process in parallel, taking into account only the global context features, not the correlation between the sentiment polarity and the local context. To address this issue, this paper proposes an Interactive Attention Encoder Network Model with Local Context Features (IAEN-LCF) for identifying aspect-level sentiment polarity. First, the word embedding and aspect term embedding are pre-trained by Bidirectional Encoder Representations from Transformers (BERT). Secondly, the attention-over-attention (AOA) module in the machine reading comprehension task is applied to the attentional encoder network, and a network model named Interactive Attention Encoder (IAEN) is proposed to extract global context features. By setting a fixed text window, the local context features are captured in a dynamic weighted manner. Finally, the performance of the proposed model is verified in three public data sets. Experimental results show that the proposed model can outperform state-of-the-art methods in aspect sentiment analysis tasks.