Automatic Feature Selection for Agenda-Based Dependency Parsing

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 794–805, Dublin, Ireland, August 23-29 2014. Automatic Feature Selection for Agenda-Based Dependency Parsing Miguel Ballesteros Natural Language Processing Group Universitat Pompeu Fabra Barcelona, Spain miguel.ballesteros@upf.edu Bernd Bohnet School of Computer Science University of Birmingham Birmingham, United Kingdom bohnetb@cs.bham.ac.uk Abstract In this paper we present an in-depth study on automatic feature selection for beam-search depen- dency parsers. The search strategy is inherited from the one implemented in MaltOptimizer, but searches in a much larger set of feature templates that could lead to a higher number of combina- tions. Our models provide results that are on par with models trained with a larger set of feature templates, and this implies that our models provide faster training and parsing times. Moreover, the results establish the state of the art for some of the languages. 1 Introduction Finding an optimal and accurate set of feature templates is crucial when training statistical parsers; in fact it is essential when building any machine learning system (Smith, 2011). In dependency parsing, the features are based on the linguistic information that is annotated within the words and the information that is being calculated during the parsing process. Researchers normally tend to include a large set of feature templates in their machine learning models, following the idea that more is always better; however some recent research on feature selection for transition-based parsing (Ballesteros, 2013; Ballesteros and Nivre, 2014) and graph-based parsing (He et al., 2013) have shown that more features are not always better, at least in the case of dependency parsing; models containing more features are always slower in parsing and training time and they do not always provide better results. This indicates that a smart feature template selection could be the key in the trade-off for finding an accurate and fast feature model for a given parsing model. On the one hand, we want a parser that should provide the best results possible, while on the other hand, we want a parser that should provide the results in the fastest way possible. For practical applications, a fast model is crucial. In this paper, we report the results of feature selection experiments that we carried out with the in- tention of obtaining accurate and faster feature models, for the transition-based Mate parser with and without graph-based completion models. The Mate parser is a beam search parser that uses a hash kernel for training, joint part-of-speech tagging, morphological tagging and dependency parsing. As a result of this research, we provide a framework that allows to find an optimal feature template set for the Mate parser (Bohnet et al., 2013). Moreover, our models provide some of the highest results ever reported for a set of treebanks. The paper is organized as follows. Section 2 describes related work including the used agenda-based dependency parser. This section depicts the feature templates that can be used by a transition-based or a graph-based parser. Section 3 describes the feature selection …

Content truncated to 500 words