Inducing Lexical Style Properties for Paraphrase and Genre Differentiation

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 218–224, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Inducing Lexical Style Properties for Paraphrase and Genre Differentiation Ellie Pavlick University of Pennsylvania epavlick@seas.upenn.edu Ani Nenkova University of Pennsylvania nenkova@seas.upenn.edu Abstract We present an intuitive and effective method for inducing style scores on words and phrases. We exploit signal in a phrase’s rate of occurrence across stylistically contrasting cor- pora, making our method simple to implement and efficient to scale. We show strong results both intrinsically, by correlation with human judgements, and extrinsically, in applications to genre analysis and paraphrasing. 1 Introduction True language understanding requires comprehend- ing not just what is said, but how it is said, yet only recently have computational approaches been applied to the subtleties of tone and style. As the expectations on language technologies grow to in- clude tailored search, context-aware inference, and analysis of author belief, an understanding of style becomes crucial. Lexical features have proven indispensable for the good performance of most applications dealing with language. Particularly, more generalized characteri- zations of the lexicon (Brown et al., 1992; Wilson et al., 2005; Feng et al., 2013; Ji and Lin, 2009; Resnik, 1995) have become key in overcoming issues with lexical sparseness and in providing practical seman- tic information for natural language processing sys- tems (Miller et al., 2004; Rutherford and Xue, 2014; Velikovich et al., 2010; Dodge et al., 2012). Most work on stylistic variation, however, has focused on larger units of text (Louis and Nenkova, 2013; Danescu-Niculescu-Mizil et al., 2012; Greene and Resnik, 2009; Xu et al., 2012) and studies of style at the lexical level have been scant. The few recent ef- forts (Brooke et al., 2010; Brooke and Hirst, 2013b; Formal/Casual Complex/Simple jesus/my gosh great/a lot 18 years/eighteen cinema/a movie respiratory/breathing a large/a big yes/yeah music/the band decade/ten years much/many things 1970s/the seventies exposure/the show foremost/first of all relative/his family megan/you there matters/the things somewhere/some place april/apr this film/that movie journal/diary full/a whole bunch the world/everybody otherwise/another thing burial/funeral father/my dad rail/the train recreation/hobby physicians/a doctor Table 1: Paraphrases with large style differences. Our method learns these distinctions automatically. Brooke and Hirst, 2013a) have been motivated by the need to categorize genre in multiple continuous dimensions and focused on applying standard meth- ods for lexical characterization via graph propaga- tion or crowdsourcing. We propose a simple and flexible method for plac- ing phrases along a style spectrum. We focus on two dimensions: formality and complexity. We eval- uate the resulting scores in terms of their correla- tion with human judgements as well as their util- ity in two tasks. First, we use the induced dimen- sions to identify stylistic shifts in paraphrase, allow- ing us to differentiate stylistic properties in the Para- phrase Database (PPDB) with high accuracy. Sec- ond, we test how well the induced scores capture differences between genres, and explore the extent to which these differences …

Content truncated to 500 words