Enforcing Subcategorization Constraints in a Parser Using Sub-parses Recombining

Proceedings of NAACL-HLT 2013, pages 239–247, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Enforcing Subcategorization Constraints in a Parser Using Sub-parses Recombining Seyed Abolghasem Mirroshandel†,⋆ Alexis Nasr† Benoˆıt Sagot⋄ †Laboratoire d’Informatique Fondamentale de Marseille- CNRS - UMR 7279 Universit´e Aix-Marseille, Marseille, France ⋄Alpage, INRIA & Universit´e Paris-Diderot, Paris, France ⋆Computer Engineering Department, Faculty of Engineering, University of Guilan, Rasht, Iran (ghasem.mirroshandel@lif.univ-mrs.fr, alexis.nasr@lif.univ-mrs.fr, benoit.sagot@inria.fr) Abstract Treebanks are not large enough to adequately model subcategorization frames of predica- tive lexemes, which is an important source of lexico-syntactic constraints for parsing. As a consequence, parsers trained on such tree- banks usually make mistakes when selecting the arguments of predicative lexemes. In this paper, we propose an original way to correct subcategorization errors by combining sub- parses of a sentence S that appear in the list of the n-best parses of S. The subcatego- rization information comes from three differ- ent resources, the first one is extracted from a treebank, the second one is computed on a large corpora and the third one is an existing syntactic lexicon. Experiments on the French Treebank showed a 15.24% reduction of er- roneous subcategorization frames (SF) selec- tions for verbs as well as a relative decrease of the error rate of 4% Labeled Accuracy Score on the state of the art parser on this treebank. 1 Introduction Automatic syntactic parsing of natural languages has witnessed many important changes in the last fifteen years. Among these changes, two have mod- ified the nature of the task itself. The first one is the availability of treebanks such as the Penn Tree- bank (Marcus et al., 1993) or the French Treebank (Abeill´e et al., 2003), which have been used in the parsing community to train stochastic parsers, such as (Collins, 1997; Petrov and Klein, 2008). Such work remained rooted in the classical language the- oretic tradition of parsing, generally based on vari- ants of generative context free grammars. The sec- ond change occurred with the use of discriminative machine learning techniques, first to rerank the out- put of a stochastic parser (Collins, 2000; Charniak and Johnson, 2005) and then in the parser itself (Rat- naparkhi, 1999; Nivre et al., 2007; McDonald et al., 2005a). Such parsers clearly depart from classical parsers in the sense that they do not rely anymore on a generative grammar: given a sentence S, all pos- sible parses for S1 are considered as possible parses of S. A parse tree is seen as a set of lexico-syntactic features which are associated to weights. The score of a parse is computed as the sum of the weights of its features. This new generation of parsers allows to reach high accuracy but possess their own limitations. We will focus in this paper on one kind of weakness of such parser which is their inability to properly take into account subcategorization frames (SF) of predicative lexemes2, an important source of lexico- syntactic constraints. The proper treatment of SF is actually confronted to two kinds of problems: …

Content truncated to 500 words