Sentence Subjectivity Detection with Weakly-Supervised Learning

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1153–1161, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Sentence Subjectivity Detection with Weakly-Supervised Learning Chenghua Lin1,2, Yulan He2 and Richard Everson1 1Department of Computer Science, University of Exeter, Exeter, EX4 4QF, UK 2Knowledge Media Institute, The Open University, Milton Keynes, MK7 6AA, UK {C.Lin,Y.He}@open.ac.uk, R.M.Everson@exeter.ac.uk Abstract This paper presents a hierarchical Bayesian model based on latent Dirichlet allocation (LDA), called subjLDA, for sentence-level subjectivity detection, which automatically identifies whether a given sentence expresses opinion or states facts. In contrast to most of the existing methods relying on either labelled corpora for classifier train- ing or linguistic pattern extraction for subjectivity classification, we view the problem as weakly-supervised generative model learning, where the only input to the model is a small set of domain independent subjectivity lexical clues. A mechanism is introduced to incor- porate the prior information about the subjectivity lexical clues into model learning by modifying the Dirichlet priors of topic-word distributions. The subjLDA model has been evaluated on the Multi-Perspective Question Answering (MPQA) dataset and promising results have been observed in the preliminary experiments. We have also explored adding neutral words as prior information for model learning. It was found that while incorporating subjectivity clues bearing positive or negative polarity can achieve a significant performance gain, the prior lexical information from neutral words is less effective. 1 Introduction Subjectivity detection seeks to identify whether the given text expresses opinions (subjective) or reports facts (objective). Such a task of distin- guishing subjective information from objective is useful for many natural language processing ap- plications. For instance, sentiment classification often assumes that the input documents are opin- ionated, and ideally only contain subjective state- ments. Document summarization systems need to summarize different perspectives and opinions. For question answering systems, extracting and presenting information of the appropriate type, i.e., opinions or facts, is imperative according to the specific question being asked (Yu and Hatzi- vassiloglou, 2003; Wiebe and Riloff, 2005; Pang and Lee, 2008). Work on sentence-level subjectivity detection is relatively sparse compared to document-level sentiment classification. Early work used a boot- strapping algorithm to learn subjective (Riloff and Wiebe, 2003) or both subjective and objec- tive (Wiebe and Riloff, 2005) expressions for sentence-level subjectivity detection. In contrast to bootstrapping, there has been some recent at- tempts exploring various n-gram features and dif- ferent level of lexical instantiation for detecting subjective utterance from conversation data (Wil- son and Raaijmakers, 2008; Raaijmakers et al., 2008; Murray and Carenini, 2009). However, the aforementioned line of work tack- led subjectivity detection either as supervised or semi-supervised learning, requiring labelled data and extensive knowledge which are expensive to acquire. On the other hand, both subjectivity and sentiment are context sensitive and in general quite domain dependent (Pang and Lee, 2008), so that classifiers trained on one domain often fail to produce satisfactory performance when shifted to new domains (Gamon et al., 2005; Blitzer et al., 2007). Moreover, user generated content …

Content truncated to 500 words