The Influence of Minimum Edit Distance on Reference Resolution

The Influence of Minimum Edit Distance on Reference Resolution Michael Strube European Media Laboratory GmbH Villa Bosch Schloß-Wolfsbrunnenweg 33 69118 Heidelberg, Germany strube@eml.villa-bosch.de Stefan Rapp Sony International (Europe) GmbH Advanced Technology Center Stuttgart Heinrich-Hertz-Straße 1 70327 Stuttgart, Germany rapp@sony.de Christoph M¨uller European Media Laboratory GmbH Villa Bosch Schloß-Wolfsbrunnenweg 33 69118 Heidelberg, Germany mueller@eml.villa-bosch.de Abstract We report on experiments in reference res- olution using a decision tree approach. We started with a standard feature set used in previous work, which led to moderate re- sults. A closer examination of the perfor- mance of the features for different forms of anaphoric expressions showed good re- sults for pronouns, moderate results for proper names, and poor results for definite noun phrases. We then included a cheap, language and domain independent feature based on the minimum edit distance be- tween strings. This feature yielded a sig- nificant improvement for data sets consist- ing of definite noun phrases and proper names, respectively. When applied to the whole data set the feature produced a smaller but still significant improvement. 1 Introduction For the automatic understanding of written or spo- ken natural language it is crucial to be able to iden- tify the entities referred to by referring expressions. The most common and thus most important types of referring expressions are pronouns and definite noun phrases (NPs). Supervised machine learning algorithms have been used for pronoun resolution (Ge et al., 1998) and for the resolution of definite NPs (Aone and Bennett, 1995; McCarthy and Lehn- ert, 1995; Soon et al., 2001). An unsupervised ap- proach to the resolution of definite NPs was applied by Cardie and Wagstaff (1999). However, though machine learning algorithms may deduce to make best use of a given set of features for a given prob- lem, it is a linguistic question and a non-trivial task to identify a set of features which describe the data sufficiently. We report on experiments in the resolution of anaphoric expressions in general, including definite noun phrases, proper names, and personal, posses- sive and demonstrative pronouns. Based on the work mentioned above we started with a feature set including NP-level and coreference-level features. Applied to the whole data set these features led only to moderate results. Since the NP form of the anaphor (i.e., whether the anaphoric expression is realized as pronoun, definite NP or proper name) ap- peared to be the most important feature, we divided the data set into several subsets based on the NP form of the anaphor. This led to the insight that the moderate performance of our system was caused by the low performance for definite NPs. We adopted a new feature based on the minimum edit distance (Wagner and Fischer, 1974) between anaphor and antecedent, which led to a significant improvement on definite NPs and proper names. When applied to the whole data set the feature yielded a smaller but still significant improvement. In this paper, we first discuss features that have been found to be relevant for the task of reference …

Content truncated to 500 words