Sentiment Lexicon Construction with Representation Learning Based on Hierarchical Sentiment Supervision

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 502–510 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Sentiment Lexicon Construction with Representation Learning Based on Hierarchical Sentiment Supervision Leyi Wang and Rui Xia∗ School of Computer Science & Engineering Nanjing University of Science & Technology, China leyiwang.cn@gmail.com, rxia@njust.edu.cn Abstract Sentiment lexicon is an important tool for identifying the sentiment polarity of word- s and texts. How to automatically con- struct sentiment lexicons has become a research topic in the field of sentimen- t analysis and opinion mining. Recently there were some attempts to employ repre- sentation learning algorithms to construct a sentiment lexicon with sentiment-aware word embedding. However, these method- s were normally trained under document- level sentiment supervision. In this paper, we develop a neural architecture to train a sentiment-aware word embedding by inte- grating the sentiment supervision at both document and word levels, to enhance the quality of word embedding as well as the sentiment lexicon. Experiments on the SemEval 2013-2016 datasets indicate that the sentiment lexicon generated by our ap- proach achieves the state-of-the-art per- formance in both supervised and unsuper- vised sentiment classification, in compari- son with several strong sentiment lexicon construction methods. 1 Introduction Sentiment lexicon is a set of words (or phrases) each of which is assigned with a sentiment polar- ity score. Sentiment lexicon plays an important role in many practical sentiment analysis and opin- ion mining tasks. There were some manually an- notated universal sentiment lexicons such as Gen- eral Inquireer (GI) and HowNet. However, due to the ubiquitous domain diversity and absence of domain prior knowledge, the automatic construc- tion technique for domain-specific sentiment lex- ∗The corresponding author of this paper. icons has become a challenging research topic in the field of sentiment analysis and opinion mining (Wang and Xia, 2016). The early work employed unsupervised learn- ing for sentiment lexicon construction. They nor- mally labelled a set of seed words at first, and then learned the polarity of each candidate word, based on either word conjunction relations (e.g., constel- lation and transition in texts) (Hatzivassiloglou and McKeown, 1997), or the word co-occurrence information (such as pointwise mutual informa- tion, PMI) (Turney, 2002), between the candidate word and the seed words. However, the unsuper- vised manner showed limited effect in sentiment prediction, and the performance greatly depends on the quality of the seed words. To fully exploit the sentiment labeling informa- tion in texts, a series of supervised learning meth- ods was further proposed to learn the sentiment lexicons. For example, Mohammad et al. (2013) proposed to construct sentiment lexicons by cal- culating PMI between the word and the distant- ly supervised sentiment labels (such as emoticon- s) in tweets and the word’s sentiment orientation (SO). The resulting lexicons obtained the best re- sults in SemEval 2013. More advanced repre- sentation learning models were also utilized, with the aim to construct the sentiment lexicons with efficient word embeddings (Tang et al., 2014a; Hamilton et …

Content truncated to 500 words