ECNU at SemEval-2017 Task 5: An Ensemble of Regression Algorithms with Effective Features for Fine-Grained Sentiment Analysis in Financial Domain

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 888–893, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics ECNU at SemEval-2017 Task 5: An Ensemble of Regression Algorithms with Effective Features for Fine-Grained Sentiment Analysis in Financial Domain Mengxiao Jiang1, Man Lan1,2∗, Yuanbin Wu1,2 1Department of Computer Science and Technology, East China Normal University, Shanghai 200062, P.R.China 2Shanghai Key Laboratory of Multidimensional Information Processing 51151201080@stu.ecnu.edu.cn, mlan,ybwu@cs.ecnu.edu.cn Abstract This paper describes our systems submitted to the Fine-Grained Sentiment Analysis on Finan- cial Microblogs and News task (i.e., Task 5) in SemEval-2017. This task includes two subtasks in microblogs and news headline domain respective- ly. To settle this problem, we extract four types of effective features, including linguistic features, sentiment lexicon features, domain-specific fea- tures and word embedding features. Then we em- ploy these features to construct models by using ensemble regression algorithms. Our submissions rank 1st and rank 5th in subtask 1 and subtask 2 respectively. 1 Introduction SemEval-2017 Task 5 is Fine-Grained Senti- ment Analysis on Financial Microblogs and News(Cortis et al., 2017), focusing on identify- ing positive (bullish; believing that the stock price will increase) and negative (bearish; believing that the stock price will decline) sentiment associated with stocks and companies from microblogs and news domains. Unlike previous sentiment analy- sis, in this task, the fine-grained sentiment analysis not only contains sentiment orientation (i.e., pos- itive or negative of the sentiment score) but also sentiment strength (i.e., the value of the sentimen- t score) attached to a particular company or stock explicitly or implicitly expressed in given texts. Given a text instance (a microblog message from Twitter or StockTwits in subtask 1, a news statement or a headline in subtask 2), the goal of participants is to predict the sentiment score for each of the stocks and companies mentioned. The sentiment score is a floating value in the range of -1 (very negative) to 1 (very positive), with 0 designating neutral sentiment. Each microblog instance contains the following 5 items: “id”, “source” (i.e., Twitter or StockTwits), “cashtag” (i.e., the company stock symbols to be predicted, e.g. “$AAPL”), “spans” and “sentiment score”. And each news headline instance contains 4 items: “id”, “company” (i.e., the company to be predict- ed), “title” and “sentiment score”. There are several differences between the span- s in subtask 1 and the title in subtask 2: (1) The spans are sentence fragments related to the cash- tag to be predicted, whereas the title is a complete sentence; (2) The spans almost regard one cashtag while the title usually contains one or more com- panies; (3) Due to (1) and (2), the spans contain less words but more effective information and less noises, which is contrary to the title. In this work, the similar method is adopted for two subtasks. We extract a series of elaborately designed features. In addition to linguistic fea- tures, sentiment lexicon features and word em- bedding features, we also extract some domain- specific features for this …

Content truncated to 500 words