Antecedent Recovery: Experiments with a Trace Tagger

Antecedent Recovery: Experiments with a Trace Tagger P´eter Dienes and Amit Dubey Department of Computational Linguistics Saarland University PO Box 15 11 50 66041 Saarbr¨ucken, Germany {dienes,adubey}@coli.uni-sb.de Abstract This paper explores the problem of find- ing non-local dependencies. First, we isolate a set of features useful for this task. Second, we develop both a two-step approach which combines a trace tagger with a state-of-the-art lexicalized parser and a one-step approach which finds non- local dependencies while parsing. We find that the former outperforms the latter be- cause it makes better use of the features we isolate. 1 Introduction Many broad-coverage statistical parsers (Charniak, 2000; Collins, 1999; Bod, 2001) are not able to give a full interpretation for sentences such as: (1) It is difficult to guess what she wants to buy. Building the semantic interpretation of this sentence requires recovering three non-local relations: (i) the object of buy is what;1 (ii) the subject of buy is she; and (iii) guess does not have a subject in the sen- tence. Three approaches have been proposed to de- tect such relations: (i) post-processing the output of a parser not designed to detect extraction sites (Johnson, 2002); (ii) integrating antecedent recov- ery into the parser (henceforth in-processing) by ei- ther enriching a syntactically simple model (Collins, 1999) or using a more powerful syntactic framework 1Collins (1999) can handle this case (Model 3). (Clark et al., 2002; Riezler et al., 2002); and (iii) de- tecting non-local dependencies as a pre-processing step before parsing (Dienes and Dubey, 2003). While the pre-processing approach is reported to give state-of-the-art performance using unlexi- calized parsers, it has not been tested using lexi- calized models. Our main claim is that that the pre-processing approach, coupled with a lexical- ized parser outperforms both state-of-the-art post- processing and in-processing. However, we show that Model 3 of Collins (1999) can be generalized to handle all types of long-distance dependencies with performance close to the pre-processing archi- tecture. A general contribution of this paper is that it gives important insights about the nature of the problem. Recovering non-local semantic relations is regarded to be a difficult problem. The successes (and fail- ures) of the simple architecture outlined here help determine what features are to be incorporated into a parser in order to improve recovery of non-local dependencies. The overall organization of the paper is as fol- lows. First, Section 2 sketches the material we use for the experiments in the paper. In Section 3, we discuss a finite-state system, a trace tagger, that de- tects extraction sites without knowledge of phrase- structure and we isolate important cues for the task. Section 4 combines the trace tagger with a parser in order to recover antecedents. Finally, in Section 5, we investigate whether and how detection of extrac- tion sites and antecedent recovery can be integrated into a lexicalized stochastic parser. Type Freq. Explanation Example NP–NP 987 controlled NP-traces Sam was seen * WH–NP 438 NP-traces of A � -movement the woman who you saw …

Content truncated to 500 words