ECNU at SemEval-2016 Task 7: An Enhanced Supervised Learning Method for Lexicon Sentiment Intensity Ranking

Proceedings of SemEval-2016, pages 491–496, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics ECNU at SemEval-2016 Task 7: An Enhanced Supervised Learning Method for Lexicon Sentiment Intensity Ranking Feixiang Wang1, Zhihua Zhang1, Man Lan1,2∗ 1Department of Computer Science and Technology, East China Normal University, Shanghai, P.R.China 2Shanghai Key Laboratory of Multidimensional Information Processing 51151201049@ecnu.cn, 51131201039@ecnu.cn, mlan@cs.ecnu.edu.cn∗ Abstract This paper describes our system submissions to task 7 in SemEval 2016, i.e., Determining Sentiment Intensity. We participated the first two subtasks in English, which are to predict the sentiment intensity of a word or a phrase in English Twitter and General English domains. To address this task, we present a supervised learning-to-rank system to predict the relevant scores, i.e., the strength associated with posi- tive sentiment, for English words or phrases. Multiple linguistic and sentiment features are adopted, e.g., Sentiment Lexicons, Sentiment Word Vectors, Word Vectors, Linguistic Fea- tures, etc. Officially released results showed that our systems rank the 1st among all sub- missions in English, which proves the effec- tiveness of the proposed method. 1 Introduction The study of sentiment analysis is increasingly drawing attention of Natural Language Processing (NLP). Many of the top performing sentiment anal- ysis systems rely on sentiment lexicon (Tan et al., 2008; Na et al., 2009; Mohammad et al., 2013). A sentiment lexicon is a list of words and phrases, such as “excellent”, “awful” and “not bad”, each of them is assigned with a positive or negative score reflecting its sentiment polarity and strength (Tang et al., 2014a). Higher scores indicate stronger sen- timent strength. However, many existing manual- ly generated sentiment lexicons consist of lexicon- s with only sentiment orientation rather than senti- ment strength. For example, the words in BL (Ding et al., 2008) are generally divided to two classes, i.e., positive and negative. Although several sentiment lexicons have assigned discrete labels for terms, i.e., strong and weak, for example, MPQA (Wiebe et al., 2005), there is no continuous real-valued scores to indicate the intensity of sentiment so far. The task of Determining Sentiment Intensity of English and Arabic Phrases intends to automatical- ly create a sentiment lexicon with real-valued scores indicating the intensity of sentiment. The purpose of this task is to test the ability of an automatic sys- tem to predict a sentiment intensity score for a word or a phrase. Phrases include negators, modals, in- tensifiers, and diminishers. Given a list of terms, the participants are required to assign appropriate s- cores between 0 and 1, to indicate their strength of association with positive sentiment. The task con- tains three subtasks (the first two are in English and the third is in Arabic) and we participated the first two subtasks in English. The first General English Sentiment Modifiers Set contains phrases formed by a word and a modifier, where a modifier can be a negator, an auxiliary verb, a degree adverb, or even a combination of those above modifiers, e.g., “would be very easy”, “did not harm”, and …

Content truncated to 500 words