Skeletons in the parser: Using a shallow parser to improve deep parsing

Skeletons in the parser: Using a shallow parser to improve deep parsing Mary Swift, James Allen, and Daniel Gildea Department of Computer Science University of Rochester Rochester, NY 14627 �swift,allen,gildea �@cs.rochester.edu Abstract We describe a simple approach for integrating shallow and deep parsing. We use phrase struc- ture bracketing obtained from the Collins parser as filters to guide deep parsing. Our exper- iments demonstrate that our technique yields substantial gains in speed along with modest improvements in accuracy. 1 Introduction The detailed linguistic analyses generated by deep parsing are an essential component of spoken dia- log systems that collaboratively perform tasks with users (e.g., (Allen et al., 2001)). For example, inter- pretation in the TRIPS collaborative dialog assistant relies on the representation produced by its parser for word sense disambiguation, constituent depen- dencies, and semantic roles such as agent, theme, goal, etc. Broad coverage unification-based deep parsers, however, unavoidably have problems meet- ing the very high accuracy and efficiency require- ments needed for real-time dialog. On the other hand, parsers based on lexicalized probabilistic con- text free grammars such those of Collins (1999) and Charniak (1997), which we call shallow parsers1, are robust and efficient, but the structural represen- tations obtained with such parsers are insufficient as input for intelligent reasoning. In addition, they are not accurate when exact match is considered as op- posed to constituent recall and precision and bracket crossing. For example, the standard Collins parser yields an exact match on only 36% on the standard test set (section 23) of the Wall Street Journal Cor- pus. In this paper we explore the question of whether preprocessing with a shallow parser can produce analyses that are good enough to help improve the speed and accuracy of deep parsing. Previous work on German (Frank et al., 2002) pursued a similar strategy and showed promising results after consid- erable effort transforming the output of the shal- 1We do not intend the “chunking” sense of shallow parsing — all our parsers return tree structures. low parser into useful guidance to the deep parser. We were interested in seeing if we could take a shallow parser off the shelf, namely the Collins parser, and use its output fairly directly to improve the performance of the TRIPS parser. It has been reported that stochastic parsers degrade in perfor- mance on domains different than what they were trained on (Hwa, 1999; Gildea, 2001), so there re- ally was an issue whether the output would be good enough. In particular, we are taking the Collins parser trained on the Wall Street Journal and ap- plying it unchanged to spontaneous human-human dialog in an emergency rescue task domain. We have found that there are islands of reliability in the results from the Collins parser that can be used to substantially improve the performance of the TRIPS parser. The remainder of the paper is organized as fol- lows. Section 2.1 provides background on the Mon- roe corpus, a set of task-oriented dialogs that is the basis …

Content truncated to 500 words