Optimizing Algorithms for Pronoun Resolution

Optimizing Algorithms for Pronoun Resolution Michael Schiehlen � Institute for Computational Linguistics University of Stuttgart Azenbergstraße 12, D-70174 Stuttgart mike@ims.uni-stuttgart.de Abstract The paper aims at a deeper understanding of sev- eral well-known algorithms and proposes ways to optimize them. It describes and discusses factors and strategies of factor interaction used in the algo- rithms. The factors used in the algorithms and the algorithms themselves are evaluated on a German corpus annotated with syntactic and coreference in- formation (Negra) (Skut et al., 1997). A common format for pronoun resolution algorithms with sev- eral open parameters is proposed, and the parameter settings optimal on the evaluation data are given. 1 Introduction In recent years, a variety of approaches to pronoun resolution have been proposed. Some of them are based on centering theory (Strube, 1998; Strube and Hahn, 1999; Tetreault, 2001), others on Machine Learning (Aone and Bennett, 1995; Ge et al., 1998; Soon et al., 2001; Ng and Cardie, 2002; Yang et al., 2003). They supplement older heuristic approaches (Hobbs, 1978; Lappin and Leass, 1994). Unfortu- nately, most of these approaches were evaluated on different corpora making different assumptions so that direct comparison is not possible. Appreciation of the new insights is quite hard. Evaluation differs not only with regard to size and genre of corpora but also along the following lines. Scope of application: Some approaches only deal with personal and possessive pronouns (centering and heuristic), while others consider coreference links in general (Soon et al., 2001; Ng and Cardie, 2002; Yang et al., 2003). A drawback of this lat- ter view is that it mixes problems on different lev- els of difficulty. It remains unclear how much of the success is due to the virtues of the approach and how much is due to the distribution of hard and easy problems in the corpus. In this paper, we will only deal with coreferential pronouns (i.e. possessive, demonstrative, and third person pronouns). � My thanks go to Melvin Wurster for help in annotation and to Ciprian Gerstenberger for discussion. Quality of linguistic input: Some proposals were evaluated on hand annotated (Strube and Hahn, 1999) or tree bank input (Ge et al., 1998; Tetreault, 2001). Other proposals provide a more realistic picture in that they work as a backend to a parser (Lappin and Leass, 1994) or noun chunker (Mitkov, 1998; Soon et al., 2001; Ng and Cardie, 2002)). In evaluation of applications presupposing parsing, it is helpful to separate errors due to parsing from in- trinsic errors. On the other hand, one would also like to gauge the end-to-end performance of a sys- tem. Thus we will provide performance figures for both ideal (hand-annotated) input and realistic (au- tomatically generated) input. Language: Most approaches were evaluated on English where large resources are available, both in terms of pre-annotated data (MUC-6 and MUC-7 data) and lexical information (WordNet). This paper deals with German. Arguably, the free word-order of German arguably leads to a clearer distinction be- tween grammatical function, surface order, and …

Content truncated to 500 words