Distributed Text Representation with Weighting Scheme Guidance for Sentiment Analysis

With rapid growth of social media, sentiment analysis has recently attracted growing attention in both academic and industrial fields. One of the most successful paradigms for sentiment analysis is to feed bag-of-words (BOW) features into classifiers. Usually, weighting schemes are required to weight raw BOW features to obtain better accuracy, where important words are assigned more weights while unimportant ones are given less weights. Another line of researches for sentiment analysis focuses on neural models, where dense features are automatically extracted from texts by neural networks. In this paper, we take advantages of techniques in both lines of researches, where weighting schemes are introduced into the neural models to guide neural networks to focus on those important words. Neural models are known for their automatic feature learning abilities, however, we discover that when suitable guidance such as weighting schemes are applied, better features can be extracted for sentiment analysis. Experimental results show that our models outperform or can compete with state-of-the-art approaches on three commonly used sentiment analysis datasets.