A Context-Based Regularization Method for Short-Text Sentiment Analysis

Sentiment analysis is an important task in natural language processing, which has promises great value to areas of interests such as business, politics and other fields. The prevalence of the internet has caused people to prefer expressing their opinion and sentiment on the Internet via methods such as tweeting on social media and commenting on products. However, the discourse of users on social media are usually short either due to limitations on post size or time limitations of the users. What emerges from these features is that the grammar used in these posts and the word meanings are flexible, which makes sentiment analysis difficult. Therefore, sentiment analysis is an important yet challenging task. In this paper, we propose a context-based regularization classification method for short text sentiment analysis. Specifically, we use contextual knowledge obtained from the data to improve performance of the sentiment classification. In this paper, the contextual knowledge includes two parts: word-sentiment knowledge and word-similarity knowledge. Moreover, we propose methods to calculate the sentiment of the words and the similarity between words on semantic level. Specifically, on the one side, we use a TRSR method based on the TextRank algorithm to rank words in each sentiment sample to determine the sentiment polarity of each word. On the other, we calculate the similarity between words according to word-embedding. In this way, we can determine the similarity between words and the sentiment polarity of a word. We then incorporate the contextual knowledge as a regularization into a supervised classification framework, which then converts into an optimization problem to train a more accurate model. Experiments on both Chinese and English datasets outperform than other baseline approaches, which demonstrates our method to be stable and effective.