Masking Polarity Words in CNN

Abstract

Sentiment Analysis (SA), also called Opinion Mining, has attracted more and more attentions in recent years. Convolutional Neural Networks (CNNs), which have a genius for extracting features, are commonly used in natural language processing (NLP) tasks. CNNs have been proven to be effective in SA tasks since sentiment is usually determined by some polarity words and phrases. In this work, we present a mask method to improve CNNs for SA tasks. We mask words according to their polarity scores and force the model to pay more attention to inapparent features. The model is supposed to extract not only the features with strong polarity but also the features with weak polarity, which is helpful for the classification. The effectiveness of the proposed mask method is demonstrated on five sentiment analysis datasets.