Boosting for Efficient Model Selection for Syntactic Parsing

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1–11, Osaka, Japan, December 11-17 2016. Boosting for Efficient Model Selection for Syntactic Parsing Rachel Bawden⋄∗ Benoˆıt Crabb´e†‡ ⋄ LIMSI, CNRS, Univ. Paris-Sud, Universite Paris-Saclay † Alpage, Univ. Paris-Diderot (Sorbonne Paris Cit´e) & Inria ‡ Institut Universitaire de France rachel.bawden@limsi.fr benoit.crabbe@linguist.univ-paris-diderot.fr Abstract We present an efficient model selection method using boosting for transition-based constituency parsing. It is designed for exploring a high-dimensional search space, defined by a large set of feature templates, as for example is typically the case when parsing morphologically rich lan- guages. Our method removes the need to manually define heuristic constraints, which are often imposed in current state-of-the-art selection methods. Our experiments for French show that the method is more efficient and is also capable of producing compact, state-of-the-art models. 1 Introduction Model selection in feature-based parsing is crucial because features define a parsing model’s capacity to predict syntactic structure. Choosing an optimal model is a trade-off between generalisation perfor- mance, compactness and parsing speed. Although too rarely mentioned, to this we should also add the speed of the selection method, which can determine how much of the search space can actually be explored. Parsing of languages other than English, and in particular morphologically rich languages, spurred on by initiatives such as the SPMRL (Statistical Parsing of Morphologically Rich Languages) shared tasks (Seddah et al., 2014), has received a heightened interest in recent years. For such languages, it is natural to want to exploit morphologically rich data to improve parsing performance. However the effect of this is an explosion in the number of possible models due to a huge number of potential features. In this paper we introduce an efficient, language-independent model selection method for transition- based constituency parsing. It is designed for model selection when faced with a large number of possible feature templates, which is typically the case for morphologically rich languages, for which we want to exploit morphological information. The method we propose uses multi-class boosting (Zhu et al., 2006) for iterative selection in constant time, using virtually no a priori constraints on the search space. We do however introduce a pre-ranking step before selection in order to guide the selection process. We provide experiments on the adaptation of boosting for model selection in the parsing of high-dimensional data, using the transition-based lexicalised constituency parser presented in (Crabb´e, 2015) and illustrating the feasibility of the method for our working language, French. Our results show that it is possible to produce high-performing, compact models much more efficiently than naive methods. The structure of the paper is as follows. We begin by describing the transition-based parser used throughout the paper (Section 2). In Section 3 we review related work, both in model selection for parsing (Section 3.1) and on the boosting algorithm used (Section 3.2) in our proposal. In Section 4 we present our adaptation of the method for parsing and in Section 5 our experiments and results. 2 Discriminative constituency …

Content truncated to 500 words