Thai Stock News Sentiment Classification using Wordpair Features

Thai Stock News Sentiment Classification using Wordpair Features Apinan Chattupan Ponrudee Netisopakul Knowledge Management and Knowledge Engineering Laboratory (KMAKE Lab) Knowledge Management and Knowledge Engineering Laboratory (KMAKE Lab) Faculty of Information Technology Faculty of Information Technology King Mongkut’s Institute of Technology Ladkrabang, Bangkok, Thailand King Mongkut’s Institute of Technology Ladkrabang, Bangkok, Thailand S7606151@kmitl.ac.th ponrudee@it.kmitl.ac.th Abstract Thai stock brokers issue daily stock news for their customers. One broker labels these news with plus, minus and zero sign to indicate the type of recommendation. This paper proposed to classify Thai stock news by extracting important texts from the news. The extracted text is in a form of a ‘wordpair’. Three wordpair sets, manual wordpairs extraction (ME), manual wordpairs addition (MA), and automate wordpairs combination (AC), are constructed and compared for their precision, recall and f-measure. Using this broker’s news as a training set and unseen stock news from other brokers as a testing set, the experiment shows that all three sets have similar results for the training set but the second and the third set have better classification results in classifying stock news from unseen brokers. Keywords: Thai stock news, sentiment classification, text classification, wordpair features. 1 Introduction Thai stock news are daily issued from many stock brokers. Thai stock news is an important source of information for stock traders to make a decision on stock trading. However, a usual Thai stock news has a long message and sometimes not easily to interpret or conclude. One stock broker makes it easier by labeling each news with plus (+), minus (-) and zero (0) sign, to indicate the type of news as positive, negative and neutral. This automatically classifies the news into three classes. In this research, we assume that ‘features’ that can be used for classifying the news must be presented as text in the news. Although this assumption could be too strong in general, for our sole purpose of investigation, our focus here is on text form of the news. Therefore, we proposed to construct a set of these ‘texts’ to be used as features in order to classify Thai stock news into three sentiments: positive, negative and neutral classes, using known sentiment news as a training set and unseen news as a testing set. Each feature is called a wordpair, since it is a pair of a keyword and a polarity word. A keyword is a word that signifies upcoming information. A polarity word is a word associated with a keyword and signifies a sentiment that related to the keyword. Following the classification from one broker, there are three sentiments: positive, negative, and neutral. However, due to the flexibility of Thai language, the order of a keyword, a polarity word and a stock symbol may not be the same in the news. That is - a keyword may come before or after a polarity word. In addition, they may come before, after or between the stock symbols they intend to recommend. PACLIC 29 188 29th Pacific Asia Conference on Language, …

Content truncated to 500 words