A Joint Segmentation and Classification Framework for Sentiment Analysis

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 477–487, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics A Joint Segmentation and Classification Framework for Sentiment Analysis Duyu Tang♮∗, Furu Wei‡, Bing Qin♮, Li Dong♯∗, Ting Liu♮, Ming Zhou‡ ♮Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology, China ‡Microsoft Research, Beijing, China ♯Beihang University, Beijing, China ♮{dytang, qinb, tliu}@ir.hit.edu.cn ‡{fuwei, mingzhou}@microsoft.com ♯donglixp@gmail.com Abstract In this paper, we propose a joint segmenta- tion and classification framework for sen- timent analysis. Existing sentiment clas- sification algorithms typically split a sen- tence as a word sequence, which does not effectively handle the inconsistent senti- ment polarity between a phrase and the words it contains, such as “not bad” and “a great deal of”. We address this issue by developing a joint segmentation and classification framework (JSC), which si- multaneously conducts sentence segmen- tation and sentence-level sentiment classi- fication. Specifically, we use a log-linear model to score each segmentation candi- date, and exploit the phrasal information of top-ranked segmentations as features to build the sentiment classifier. A marginal log-likelihood objective function is de- vised for the segmentation model, which is optimized for enhancing the sentiment classification performance. The joint mod- el is trained only based on the annotat- ed sentiment polarity of sentences, with- out any segmentation annotations. Experi- ments on a benchmark Twitter sentimen- t classification dataset in SemEval 2013 show that, our joint model performs com- parably with the state-of-the-art methods. 1 Introduction Sentiment classification, which classifies the senti- ment polarity of a sentence (or document) as posi- tive or negative, is a major research direction in the field of sentiment analysis (Pang and Lee, 2008; Liu, 2012; Feldman, 2013). Majority of existing approaches follow Pang et al. (2002) and treat sen- ∗ This work was partly done when the first and fourth authors were visiting Microsoft Research. timent classification as a special case of text cate- gorization task. Under this perspective, previous studies typically use pipelined methods with two steps. They first produce sentence segmentation- s with separate text analyzers (Choi and Cardie, 2008; Nakagawa et al., 2010; Socher et al., 2013b) or bag-of-words (Paltoglou and Thelwall, 2010; Maas et al., 2011). Then, feature learning and sen- timent classification algorithms take the segmenta- tion results as inputs to build the sentiment classi- fier (Socher et al., 2011; Kalchbrenner et al., 2014; Dong et al., 2014). The major disadvantage of a pipelined method is the problem of error propagation, since sen- tence segmentation errors cannot be corrected by the sentiment classification model. A typical kind of error is caused by the polarity inconsistency be- tween a phrase and the words it contains, such as ⟨not bad, bad⟩ and ⟨a great deal of, great⟩. The segmentations based on bag-of-words or syn- tactic chunkers are not effective enough to han- dle the polarity inconsistency phenomenons. The reason lies in that bag-of-words segmentations re- gard each word as a separate unit, which losses the …

Content truncated to 500 words