Abstract: Text sentiment mining is widely used in empirical studies for stock prices forecasting or trend analysis. Existing researches tend to focus merely on mining the full texts of financial news with an underlying assumption that all contents are related to its topic and all sections count equally. Clearly this is unduly restrictive due to the presence of irregular text format and structural confusion of internet news. This paper proposes and studies stock trend prediction models with text summarization and sentiment mining. We use Seq2Seq method to summarize news texts automatically, then mine their sentiment to improve the performance of models, and finally incorporate these sentiment values as additional features into the machine learning of stock prediction. We conduct experiments to compare two strategies, one with sentiment values of full texts and the other with sentiment value of text summary. The results show that a) their curves of emotional fluctuation are the same in most cases, but the one with text summary has smaller range and is more stable; b) the one with text summary is more sensitive to negative emotional words; c) the one with text summary, in the stock market trend prediction, obviously dominates the other.

Key words: stock price trend, automatic text summarization, Seq2Seq, sentiment analysis