Building Large-Scale Twitter-Specific Sentiment Lexicon : A Representation Learning Approach

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 172–182, Dublin, Ireland, August 23-29 2014. Building Large-Scale Twitter-Specific Sentiment Lexicon : A Representation Learning Approach Duyu Tang♮∗, Furu Wei‡, Bing Qin♮†, Ming Zhou‡, Ting Liu♮ ♮Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology, China ‡Microsoft Research, Beijing, China {dytang, qinb, tliu}@ir.hit.edu.cn {fuwei, mingzhou}@microsoft.com Abstract In this paper, we propose to build large-scale sentiment lexicon from Twitter with a representation learning approach. We cast sentiment lexicon learning as a phrase-level sentiment classification task. The challenges are developing effective feature representation of phrases and obtaining training data with minor manual annotations for building the sentiment classifier. Specifical- ly, we develop a dedicated neural architecture and integrate the sentiment information of tex- t (e.g. sentences or tweets) into its hybrid loss function for learning sentiment-specific phrase embedding (SSPE). The neural network is trained from massive tweets collected with positive and negative emoticons, without any manual annotation. Furthermore, we introduce the Urban Dictionary to expand a small number of sentiment seeds to obtain more training data for building the phrase-level sentiment classifier. We evaluate our sentiment lexicon (TS-Lex) by applying it in a supervised learning framework for Twitter sentiment classification. Experiment results on the benchmark dataset of SemEval 2013 show that, TS-Lex yields better performance than previously introduced sentiment lexicons. 1 Introduction A sentiment lexicon is a list of words and phrases, such as “excellent”, “awful” and “not bad”, each of which is assigned with a positive or negative score reflecting its sentiment polarity and strength. Sentiment lexicon is crucial for sentiment analysis (or opining mining) as it provides rich sentiment in- formation and forms the foundation of many sentiment analysis systems (Pang and Lee, 2008; Liu, 2012; Feldman, 2013). Existing sentiment lexicon learning algorithms mostly utilize propagation methods to estimate the sentiment score of each phrase. These methods typically employ parsing results, syntac- tic contexts or linguistic information from thesaurus (e.g. WordNet) to calculate the similarity between phrases. For example, Baccianella et al. (2010) use the glosses information from WordNet; Velikovich et al. (2010) represent each phrase with its context words from the web documents; Qiu et al. (2011) exploit the dependency relations between sentiment words and aspect words. However, parsing information and the linguistic information from WordNet are not suitable for constructing large-scale sentiment lexicon from Twitter. The reason lies in that WordNet cannot well cover the colloquial expressions in tweets, and it is hard to have reliable tweet parsers due to the informal language style. In this paper, we propose to build large-scale sentiment lexicon from Twitter with a representation learning approach, as illustrated in Figure 1. We cast sentiment lexicon learning as a phrase-level classi- fication task. Our method contains two part: (1) a representation learning algorithm to effectively learn the continuous representation of phrases, which are used as features for phrase-level sentiment classifica- tion, (2) a seed expansion algorithm that enlarge a small list of sentiment seeds to collect training …

Content truncated to 500 words