BiLSTM with Multi-Polarity Orthogonal Attention for Implicit Sentiment Analysis

Abstract

Sentiment analysis has been a popular field in natural language processing. Sentiments can be expressed explicitly or implicitly. Most current studies on sentiment analysis focus on the identification of explicit sentiments. However, implicit sentiment analysis has become one of the most difficult tasks in sentiment analysis due to the absence of explicit sentiment words. In this article, we propose a BiLSTM model with multi-polarity orthogonal attention for implicit sentiment analysis. Compared to the traditional single attention model, the difference between the words and the sentiment orientation can be identified by using multi-polarity attention. This difference can be regarded as a significant feature for implicit sentiment analysis. Moreover, an orthogonal restriction mechanism is adopted to ensure that the discriminatory performance can be maintained during optimization. The experimental results on the SMP2019 implicit sentiment analysis dataset and two explicit sentiment analysis datasets demonstrate that our model more accurately captures the characteristic differences among sentiment polarities.