Topic-wise, Sentiment-wise, or Otherwise? Identifying the Hidden Dimension for Unsupervised Text Classification

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 580–589, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Topic-wise, Sentiment-wise, or Otherwise? Identifying the Hidden Dimension for Unsupervised Text Classification Sajib Dasgupta and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {sajib,vince}@hlt.utdallas.edu Abstract While traditional work on text clustering has largely focused on grouping docu- ments by topic, it is conceivable that a user may want to cluster documents along other dimensions, such as the author’s mood, gender, age, or sentiment. Without know- ing the user’s intention, a clustering al- gorithm will only group documents along the most prominent dimension, which may not be the one the user desires. To ad- dress this problem, we propose a novel way of incorporating user feedback into a clustering algorithm, which allows a user to easily specify the dimension along which she wants the data points to be clus- tered via inspecting only a small number of words. This distinguishes our method from existing ones, which typically re- quire a large amount of effort on the part of humans in the form of document an- notation or interactive construction of the feature space. We demonstrate the viabil- ity of our method on several challenging sentiment datasets. 1 Introduction Text clustering is one of the most important appli- cations in Natural Language Processing (NLP). A common approach to this problem consists of (1) computing the similarity between each pair of doc- uments, each of which is typically represented as a bag of words; and (2) using an unsupervised clus- tering algorithm to partition the documents. The majority of existing work on text clustering has focused on topic-based clustering, where high ac- curacies can be achieved even for datasets with a large number of classes (e.g., 20 Newsgroups). On the other hand, there has been relatively lit- tle work on sentiment-based clustering and the re- lated task of unsupervised polarity classification, where the goal is to cluster (or classify) a set of documents (e.g., reviews) according to the po- larity (e.g., “thumbs up” or “thumbs down”) ex- pressed by the author in an unsupervised man- ner. Despite the large amount of recent work on sentiment analysis and opinion mining, much of it has focused on supervised methods (e.g., Pang et al. (2002), Kim and Hovy (2004), Mullen and Collier (2004)). One weakness of these existing supervised polarity classification systems is that they are typically domain- and language-specific. Hence, when given a new domain or language, one needs to go through the expensive process of collecting a large amount of annotated data in or- der to train a high-performance polarity classifier. Some recent attempts have been made to leverage existing sentiment corpora or lexica to automati- cally create annotated resources for new domains or languages. However, such methods require the existence of either a parallel corpus/machine translation engine for projecting/translating anno- tations/lexica from a resource-rich language to the target language (Banea et al., 2008; Wan, …

Content truncated to 500 words