Learning Stock Market Sentiment Lexicon and Sentiment-Oriented Word Vector from StockTwits

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 301–310, Vancouver, Canada, August 3 - August 4, 2017. c⃝2017 Association for Computational Linguistics Learning Stock Market Sentiment Lexicon and Sentiment-Oriented Word Vector from StockTwits Quanzhi Li, Sameena Shah Research and Development Thomson Reuters 3 Times Square, New York, NY 10036 {quanzhi.li, sameena.shah@thomsonreuters.com} Abstract Previous studies have shown that investor sentiment indicators can predict stock market change. A domain-specific senti- ment lexicon and sentiment-oriented word embedding model would help the senti- ment analysis in financial domain and stock market. In this paper, we present a new approach to learning stock market lexicon from StockTwits, a popular finan- cial social network for investors to share ideas. It learns word polarity by predicting message sentiment, using a neural net- work. The sentiment-oriented word embeddings are learned from tens of mil- lions of StockTwits posts, and this is the first study presenting sentiment-oriented word embeddings for stock market. The experiments of predicting investor senti- ment show that our lexicon outperformed other lexicons built by the state-of-the-art methods, and the sentiment-oriented word vector was much better than the general word embeddings. 1 Introduction Social media has provided a rich opinion content that is valuable for diverse decision-making pro- cesses (Montoyo et al., 2012; Oliveira 2016), and sentiment analysis is being increasingly used to predict stock market variables (Antweiler and Frank 2014; Yu eta l., 2013; Schumaker et al., 2012). In particular, social media messages are a useful source for supporting stock market deci- sions (Bollen et al., 2011; Oliveira et al., 2013). Users of social media, such as StockTwits and Twitter, post very frequently, and this makes the real-time assessment possible, which can be exploited during the trading day. The two im- portant sentiment data that can help sentiment analysis greatly are sentiment lexicons and word embeddings learned from large amount of data. Word embedding (word vector) has been used in many NLP tasks and noticeably improved their performance (Socher et al., 2013; Tang et al., 2014b; Vo and Zhang, 2015; Li et al., 2017). However, there has been little effort in construct- ing sentiment lexicon for financial domain and stock market, and in using social media as the data source. Many terms in financial market have different meanings, especially sentiment polarity, from that in other domains or sources, such as the general news articles and Twitter. For exam- ple, terms long, short, put and call have special meanings in stock market. Another example is the term underestimate, which is a negative term in general, but it can suggest an opportunity to buy equities when is used in stock market mes- sages. Domain independent lexicons or general word embedding model may not perform well in financial domain. Therefore, it is necessary and important to built sentiment lexicons and word embeddings specifically for stock market. The automatic lexicon creation approaches in previous studies are mainly based on statistic measures. There are few studies exploiting ma- chine learning models (Tang et al, 2014a; …

Content truncated to 500 words