Trace Prediction and Recovery with Unlexicalized PCFGs and Slash Features

Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 177–184, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Trace Prediction and Recovery With Unlexicalized PCFGs and Slash Features Helmut Schmid IMS, University of Stuttgart schmid@ims.uni-stuttgart.de Abstract This paper describes a parser which gen- erates parse trees with empty elements in which traces and fillers are co-indexed. The parser is an unlexicalized PCFG parser which is guaranteed to return the most probable parse. The grammar is extracted from a version of the PENN treebank which was automatically anno- tated with features in the style of Klein and Manning (2003). The annotation in- cludes GPSG-style slash features which link traces and fillers, and other features which improve the general parsing accu- racy. In an evaluation on the PENN tree- bank (Marcus et al., 1993), the parser outperformed other unlexicalized PCFG parsers in terms of labeled bracketing f- score. Its results for the empty cate- gory prediction task and the trace-filler co- indexation task exceed all previously re- ported results with 84.1% and 77.4% f- score, respectively. 1 Introduction Empty categories (also called null elements) are used in the annotation of the PENN treebank (Mar- cus et al., 1993) in order to represent syntactic phenomena like constituent movement (e.g. wh- extraction), discontinuous constituents, and miss- ing elements (PRO elements, empty complemen- tizers and relative pronouns). Moved constituents are co-indexed with a trace which is located at the position where the moved constituent is to be interpreted. Figure 1 shows an example of con- stituent movement in a relative clause. Empty categories provide important informa- tion for the semantic interpretation, in particular NP NP NNS things SBAR WHPP-1 IN of WHNP WDT which S NP-SBJ PRP they VP VBP are ADJP-PRD JJ unaware PP -NONE- *T*-1 Figure 1: Co-indexation of traces and fillers for determining the predicate-argument structure of a sentence. However, most broad-coverage sta- tistical parsers (Collins, 1997; Charniak, 2000, and others) which are trained on the PENN tree- bank generate parse trees without empty cate- gories. In order to augment such parsers with empty category prediction, three rather different strategies have been proposed: (i) pre-processing of the input sentence with a tagger which inserts empty categories into the input string of the parser (Dienes and Dubey, 2003b; Dienes and Dubey, 2003a). The parser treats the empty elements like normal input tokens. (ii) post-processing of the parse trees with a pattern matcher which adds empty categories after parsing (Johnson, 2001; Campbell, 2004; Levy and Manning, 2004) (iii) in-processing of the empty categories with a slash percolation mechanism (Dienes and Dubey, 2003b; Dienes and Dubey, 2003a). The empty el- ements are here generated by the grammar. Good results have been obtained with all three approaches, but (Dienes and Dubey, 2003b) re- ported that in their experiments, the in-processing of the empty categories only worked with lexi- calized parsing. They explain that their unlex- 177 icalized PCFG parser produced poor results be- cause the beam search strategy …

Content truncated to 500 words