Language model based Chinese financial news sentiment classification

This paper address the problem of identifying the sentiment polarity in financial news articles about a public company having potential effect on the future price of the company's stock. The problem is challenging due to the lack of reliable labeled training data and effective classification method. A feasible corpus building strategy is proposed and stock reviews are used for training, since the news polarity prediction is similar to the process of stock analyst drawing their conclusion by weighting the major event pros and cons of the company. The reviews can be annotated automatically by the grade given by the analyst. In addition, the consequent experiments also confirm it. Furthermore, we examine the effectiveness of using language modeling approaches to solve the sentiment classification of Chinese financial news articles. Two different approaches based on language model are employed and their comparisons with SVM and Naive Bayes are also performed in our research. The experiment results justify the effectiveness and robustness of the proposed language model approaches, which perform better than the approaches based on traditional machine learning techniques.