ConnotationWordNet: Learning Connotation over the Word+Sense Network

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 1544–1554, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics ConnotationWordNet: Learning Connotation over the Word+Sense Network Jun Seok Kang Song Feng Leman Akoglu Yejin Choi Department of Computer Science Stony Brook University Stony Brook, NY 11794-4400 junkang, songfeng, leman, ychoi@cs.stonybrook.edu Abstract We introduce ConnotationWordNet, a con- notation lexicon over the network of words in conjunction with senses. We formulate the lexicon induction problem as collec- tive inference over pairwise-Markov Ran- dom Fields, and present a loopy belief propagation algorithm for inference. The key aspect of our method is that it is the first unified approach that assigns the polarity of both word- and sense-level connotations, exploiting the innate bipar- tite graph structure encoded in WordNet. We present comprehensive evaluation to demonstrate the quality and utility of the resulting lexicon in comparison to existing connotation and sentiment lexicons. 1 Introduction We introduce ConnotationWordNet, a connotation lexicon over the network of words in conjunction with senses, as defined in WordNet. A connotation lexicon, as introduced first by Feng et al. (2011), aims to encompass subtle shades of sentiment a word may conjure, even for seemingly objective words such as “sculpture”, “Ph.D.”, “rosettes”. Understanding the rich and complex layers of con- notation remains to be a challenging task. As a starting point, we study a more feasible task of learning the polarity of connotation. For non-polysemous words, which constitute a significant portion of English vocabulary, learning the general connotation at the word-level (rather than at the sense-level) would be a natural oper- ational choice. However, for polysemous words, which correspond to most frequently used words, it would be an overly crude assumption that the same connotative polarity should be assigned for all senses of a given word. For example, consider “abound”, for which lexicographers of WordNet prescribe two different senses: • (v) abound: (be abundant of plentiful; exist in large quantities) • (v) abound, burst, bristle: (be in a state of movement or action) “The room abounded with screaming children”; “The garden bris- tled with toddlers” For the first sense, which is the most commonly used sense for “abound”, the general overtone of the connotation would seem positive. That is, al- though one can use this sense in both positive and negative contexts, this sense of “abound” seems to collocate more often with items that are good to be abundant (e.g., “resources”), than unfortunate items being abundant (e.g., “complaints”). However, as for the second sense, for which “burst” and “bristle” can be used interchangeably with respect to this particular sense,1 the general overtone is slightly more negative with a touch of unpleasantness, or at least not as positive as that of the first sense. Especially if we look up the Word- Net entry for “bristle”, there are noticeably more negatively connotative words involved in its gloss and examples. This word sense issue has been a universal chal- lenge for a range of Natural Language Processing applications, including …

Content truncated to 500 words