Sentiment Lexicon Expansion Based on Neural PU Learning, Double Dictionary Lookup, and Polarity Association

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 553–563 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Sentiment Lexicon Expansion Based on Neural PU Learning, Double Dictionary Lookup, and Polarity Association Yasheng Wang1, Yang Zhang1, Bing Liu2 1Noah’s Ark Lab, Huawei Technologies 1{wangyasheng, zhangyang86}@huawei.com 2Department of Computer Science, University of Illinois at Chicago, USA 2liub@uic.com Abstract Although many sentiment lexicons in dif- ferent languages exist, most are not com- prehensive. In a recent sentiment analysis application, we used a large Chinese sen- timent lexicon and found that it missed a large number of sentiment words used in social media. This prompted us to make a new attempt to study sentiment lexi- con expansion. This paper first formu- lates the problem as a PU learning prob- lem. It then proposes a new PU learning method suitable for the problem based on a neural network. The results are further enhanced with a new dictionary lookup technique and a novel polarity classifica- tion algorithm. Experimental results show that the proposed approach greatly outper- forms baseline methods. 1 Introduction Sentiment lexicons contain words (such as good, beautiful, bad, and awful) that convey positive or negative sentiments. They are instrumental for many sentiment analysis applications. So far many algorithms have been proposed to generate such lexicons (Liu, 2012). These algorithms are either dictionary-based or corpus-based. In the dictionary-based approach, one exploits synonym and antonym relations in the dictionary to boot- strap a given seed set of sentiment words (Hu and Liu, 2004; Kim and Hovy, 2004; Kamps et al., 2004), or learns a classifier to classify the gloss of each word in the dictionary (Esuli and Sebas- tiani, 2005). The corpus-based approach uses var- ious linguistic rules or patterns (Hatzivassiloglou and McKeown, 1997; Kanayama and Nasukawa, 2006; Qiu et al., 2011; Tang et al., 2014). We will discuss these and other existing methods in the re- lated work section. Despite many existing studies, the problem is far from being solved. In a recent application, we used a popular Chinese sentiment lexicon for sentiment classification of Weibo posts (similar to Twitter), and found that it missed a large num- ber of sentiment words. As the lexicon was com- piled using formal text such as news, it misses a large number of sentiment words used in social media. Due to the informal nature, many “low class” words are used in social media but seldom used in formal media. New words are also created constantly. Note that new words in Chinese are easier to create from individual characters than in other languages. Thus many of these words are not in the dictionary. All these prompted us to make a new attempt to study sentiment lexicon expansion. In this paper, we solve the problem in two steps: (1) identify sentiment words from a given corpus, and (2) classify their polarity. We formulate Step 1 as a PU learning problem (learning from positive unlabeled examples). To our knowledge, this is the first such …

Content truncated to 500 words