Applying Co-Training to Reference Resolution

Applying Co-Training to Reference Resolution Christoph M¨uller European Media Laboratory GmbH Villa Bosch Schloß-Wolfsbrunnenweg 33 69118 Heidelberg, Germany mueller@eml.villa-bosch.de Stefan Rapp Sony International (Europe) GmbH Advanced Technology Center Stuttgart Heinrich-Hertz-Straße 1 70327 Stuttgart, Germany rapp@sony.de Michael Strube European Media Laboratory GmbH Villa Bosch Schloß-Wolfsbrunnenweg 33 69118 Heidelberg, Germany strube@eml.villa-bosch.de Abstract In this paper, we investigate the practical applicability of Co-Training for the task of building a classifier for reference reso- lution. We are concerned with the ques- tion if Co-Training can significantly re- duce the amount of manual labeling work and still produce a classifier with an ac- ceptable performance. 1 Introduction A major obstacle for natural language processing systems which analyze natural language texts or utterances is the need to identify the entities re- ferred to by means of referring expressions. Among referring expressions, pronouns and definite noun phrases (NPs) are the most prominent. Supervised machine learning algorithms were used for pronoun resolution with good results (Ge et al., 1998), and for definite NPs with fairly good re- sults (Aone and Bennett, 1995; McCarthy and Lehn- ert, 1995; Soon et al., 2001). However, the defi- ciency of supervised machine learning approaches is the need for an unknown amount of annotated train- ing data for optimal performance. So, researchers in NLP began to experiment with weakly supervised machine learning algorithms such as Co-Training (Blum and Mitchell, 1998). Among others Co-Training was applied to document classification (Blum and Mitchell, 1998), named- entity recognition (Collins and Singer, 1999), noun phrase bracketing (Pierce and Cardie, 2001), and statistical parsing (Sarkar, 2001). In this paper we apply Co-Training to the problem of reference reso- lution in German texts from the tourism domain in order to provide answers to the following questions: � Does Co-Training work at all for this task (when compared to conventional C4.5 decision tree learning)? � How much labeled training data is required for achieving a reasonable performance? First, we discuss features that have been found to be relevant for the task of reference resolution, and describe the feature set that we are using (Section 2). Then we briefly introduce the Co-Training paradigm (Section 3), which is followed by a description of the corpus we use, the corpus annotation, and the way we prepared the data for using a binary classifier in the Co-Training algorithm (Section 4). In Section 5 we specify the experimental setup and report on the results. 2 Features for Reference Resolution 2.1 Previous Work Driven by the necessity to provide robust systems for the MUC system evaluations, researchers began to look for those features which were particular im- portant for the task of reference resolution. While most features for pronoun resolution have been de- scribed in the literature for decades, researchers only recently began to look for robust and cheap features, i.e., those which perform well over several domains and can be annotated (semi-) automatically. Also, the relative quantitative contribution of each of these features came into focus only after the advent of Computational Linguistics (ACL), Philadelphia, …

Content truncated to 500 words