A Discriminative Learning Model for Coordinate Conjunctions

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 610–619, Prague, June 2007. c⃝2007 Association for Computational Linguistics A Discriminative Learning Model for Coordinate Conjunctions Masashi Shimbo∗ Graduate School of Information Science Nara Institute of Science and Technology Ikoma, Nara 630-0192, Japan shimbo@is.naist.jp Kazuo Hara∗ Graduate School of Information Science Nara Institute of Science and Technology Ikoma, Nara 630-0192, Japan kazuo-h@is.naist.jp Abstract We propose a sequence-alignment based method for detecting and disambiguating co- ordinate conjunctions. In this method, av- eraged perceptron learning is used to adapt the substitution matrix to the training data drawn from the target language and domain. To reduce the cost of training data con- struction, our method accepts training exam- ples in which complete word-by-word align- ment labels are missing, but instead only the boundaries of coordinated conjuncts are marked. We report promising empirical re- sults in detecting and disambiguating coor- dinated noun phrases in the GENIA corpus, despite a relatively small number of train- ing examples and minimal features are em- ployed. 1 Introduction Coordination, along with prepositional phrase at- tachment, is a major source of syntactic ambiguity in natural language. Although only a small number of previous studies in natural language processing have dealt with coordinations, this does not mean disambiguating coordinations is easy and negligible; it still remains one of the difficulties for state-of-the- art parsers. in Charniak and Johnson’s recent work (Charniak and Johnson, 2005), for instance, two of the features incorporated in their parse reranker are aimed specifically at resolving coordination ambi- guities. Previous work on coordinations includes (Agar- wal and Boggess, 1992; Chantree et al., 2005; Kuro- ∗Equal contribution. hashi and Nagao, 1994; Nakov and Hearst, 2005; Okumura and Muraki, 1994; Resnik, 1999). Ear- lier studies (Agarwal and Boggess, 1992; Okumura and Muraki, 1994) attempted to find heuristic rules to disambiguate coordinations. More recent re- search are concerned with capturing structural sim- ilarity between conjuncts using thesauri and cor- pora (Chantree et al., 2005), or web-based statistics (Nakov and Hearst, 2005). We identify three problems associated with the previous work. 1. Most of these studies evaluate the proposed heuristics against restricted forms of conjunc- tions. In some cases, they only deal with co- ordinations with exactly two conjuncts, leaving the generality of these heuristics unclear. 2. Most of these studies assume that the bound- aries of coordinations are known in advance, which, in our opinion, is impractical. 3. The proposed heuristics and statistics capture many different aspects of coordination. How- ever, it is not clear how they interact and how they can be combined. To address these problems, we propose a new framework for detecting and disambiguating coor- dinate conjunctions. Being a discriminative learning model, it can incorporate a large number of overlap- ping features encoding various heuristics for coordi- nation disambiguation. It thus provides a test bed for examining combined use of the proposed heuristics as well as new ones. As the weight on each feature is automatically tuned …

Content truncated to 500 words