Generative Modeling of Coordination by Factoring Parallelism and Selectional Preferences

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 456–464, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Generative Modeling of Coordination by Factoring Parallelism and Selectional Preferences Daisuke Kawahara and Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-Honmachi, Sakyo-ku, Kyoto, 606-8501, Japan {dk, kuro}@i.kyoto-u.ac.jp Abstract We present a unified generative model of coordination that considers parallelism of conjuncts and selectional preferences. Parallelism of conjuncts, which frequently characterizes coordinate structures, modeled as a synchronized generationis process in the generative parser. Selec- tional preferences learned from a large web corpus provide an important clue for resolving the ambiguities of coordinate structures. Our experiments of Japanese dependency parsing indicate the effective- ness of our approach, particularly in the domains of newspapers and patents. 1 Introduction Coordinate structures are a potential source of syn- tactic ambiguity in natural language. Although many methods have been proposed to resolve the ambiguities of coordinate structures, coordination disambiguation still remains a difficult problem for state-of-the-art parsers. Previous studies on coordination disambiguation used two kinds of clues: • parallelism of conjuncts, and • selectional preferences. Syntactic, lexical and semantic parallelism of conjuncts is frequently observed in coordinate structures. For example, Dubey et al. (2005) em- pirically confirmed syntactic parallelism in co- ordinate structures. This clue was modeled by string matching, part-of-speech matching, num- ber agreement, semantic similarities, and so forth (Agarwal and Boggess, 1992; Kurohashi and Na- gao, 1994; Resnik, 1999; Chantree et al., 2005; Buyko and Hahn, 2008). For instance, consider the following example: (1) eat Caesar salad and Italian pasta We can observe lexical or semantic parallelism be- tween salad and pasta, which can be automatically detected via a thesaurus or distributional similar- ity. In addition, syntactic parallelism can be ob- served; each conjunct has a modifier Caesar and Italian, respectively. These types of parallelism contribute to identifying the coordinate structure that conjoins Caesar salad and Italian pasta. The other clue is selectional preferences, such as eat in the above example. Since eat is likely to have salad and pasta as its objects, it is plausi- ble that salad and pasta are coordinated. Such se- lectional preferences of predicates are thought to support the construction of coordinate structures, and were used in Japanese dependency parsing by Kawahara and Kurohashi (2008). Selectional pref- erences of nouns (noun-noun modifications) were used by Resnik (1999), Nakov and Hearst (2005) and Kawahara and Kurohashi (2008). For exam- ple, let us see the following examples: (2) a. mail and securities fraud b. corn and peanut butter In (2a), the coordination of mail and securities is guided by the estimation that mail fraud is a salient compound nominal phrase. In (2b), on the con- trary, the coordinate structure that conjoins corn and peanut butter is led because corn butter is not a familiar concept. Each clue has been empirically proven to be ef- fective for coordination disambiguation. However, a unified approach that combines both clues has not been explored comprehensively. In this pa- per, we propose …

Content truncated to 500 words