An LR-inspired generalized lexicalized phrase structure parser

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 541–552, Dublin, Ireland, August 23-29 2014. An LR-inspired generalized lexicalized phrase structure parser Benoit Crabbé ALPAGE / INRIA - Université Paris Diderot Place Paul Ricoeur F-75013 Paris benoit.crabbe@univ-paris-diderot.fr Abstract The paper introduces an LR-based algorithm for efficient phrase structure parsing of morpholog- ically rich languages. The algorithm generalizes lexicalized parsing (Collins, 2003) by allowing a structured representation of the lexical items. Together with a discriminative weighting com- ponent (Collins, 2002), we show that this representation allows us to achieve state of the art accurracy results on a morphologically rich language such as French while achieving more effi- cient parsing times than the state of the art parsers on the French data set. A comparison with English, a lexically poor language, is also provided. 1 Introduction The paper provides a phrase structure parsing algorithm inspired by LR (Knuth, 1965), GLR (Tomita, 1988) and the recent developments of (Huang and Sagae, 2010) for dependency grammar. The parsing algorithm comes with a discriminative weighting framework inspired by (Collins, 2002). Although dis- criminative phrase structure parsing has been shown to be challenging when it comes to efficiency issues (Turian and Melamed, 2006; Finkel et al., 2008), we use here several approximations that make the framework not only tractable but also efficient and accurate on a lexically rich language such as French. Despite the successes of dependency grammar, we are interested in phrase structure grammar since it naturally allows to support compositional semantic representations as recently highlighted by (Socher et al., 2012). It remains that most phrase structure parsers have been designed in priority for modelling lexically poor languages such as English or Chinese (Collins, 2003; Charniak, 2000; Zhu et al., 2013). Although highly accurate multilingual parsers exist (Petrov et al., 2006), they remain relatively both slow for wide coverage purposes and their inner formal structure is not designed to handle naturally morphological information. We assume that parsing lexically rich languages benefits from taking into account the structured mor- phological information that can be extracted from lexical forms. Using French as a case study we show that we can reach both parsing efficiency with an approximative inference method and we can get a state of the art accurracy by generalizing lexicalized parsing to handle feature structure-based word represen- tations. Our proposal also differs theoretically from related ones (Sagae and Lavie, 2006; Zhang and Clark, 2011; Zhu et al., 2013) by explicitly using an LR automaton. The explicit introduction of the LR automaton allows us to establish a formal difference between shift reduce phrase structure parsing and shift reduce dependency parsing. It further provides some insights on the nature of the grammar underlying many contemporary parsers. The paper is organized as follows. First, section 2, we set up a formal framework for describing weighted phrase structure parsing as a 2-LCFG (Nederhof and Satta, 2010). Observing that the tree structures are actually constrained in practice we formulate in section 3 an LR automaton …

Content truncated to 500 words