2004 ASurveyOfMLForRefResolutionInText

• (Olsson, 2004) ⇒ Fredrik Olsson. (2004). “A Survey of Machine Learning for Reference Resolution in Textual Discourse.” Technical Report, Swedish Institute of Computer Science.

Subject Headings: Mention Coreference Resolution Task, Mention Coreference Resolution Algorithm, Anaphora Resolution Algorithm, Machine Learning Algorithm, Survey.

Quotes

Abstract

Machine learning methods have been successfully applied to a number of language technology tasks. The purpose of this report is to investigate the efforts made in applying machine learning for handling the phenomena of anaphoric reference and coreference in text.

1. Introduction

Machine learning methods have been successfully applied to a number of language technology tasks. Previous work has to a large extent been focused on intrasentential phenomenon, such as part of speech tagging, phrase chunking and named entity recognition. The purpose of this report is to investigate the efforts made in applying machine learning methods for handling the inter-sentential linguistic phenomenon of anaphoric reference and coreference in text. Due to the amount of literature available on the subject, this report will not present a paper-by-paper survey, but rather try to synthesise the information available in the sources and line up general observations. When necessary, individual approaches will be examined and disseminated. On a general note, Mitkov (2002) provides a point of departure for the computational treatment of anaphora resolution.

The term reference resolution in the title of this report is deliberately somewhat vague as it covers both coreference and anaphora. While coreference is defined as a relation holding between noun phrases if they refer to the same entity (see e.g. (Hirschman and Chinchor, 1997)), anaphora is understood as the presupposition of something that has gone before and that points back to some previous item (Halliday and Hasan, 1976). Thus, there is a difference between anaphora and coreference, and, as van Deemter and Kibble (2000) point out; coreference is an equivalence relation, while anaphora is irreexive, non-symmetrical and non-transitive relation. This imply that the phenomenon of anaphora is sensitive to context, while coreference is not.

A number of approaches has been proposed for English noun phrase reference resolution, e.g., McCarthy and Lehnert (1995), Cardie and Wagsta (1999), Soon et al (2001), Harabagiu et al (2001), Preiss (2002), Ng and Cardie (2002c; 2002b; 2002a; 2003a; 2003b), Yang et al (2003), as well as for Japanese noun phrases, e.g, Lida et al (2003), and German, e.g, Strube et al (2002), Muller et al (2002). Some attempts at solving English [1 The related concept of cataphora is rarely dealt with; in the present survey, it was found briey mentioned by Stuckardt (2002) and by Evans (2001).] anaphora is reported by, e.g, Connolly et al (1997), and Ge et al (1998), while Evans (2001) report on experiments for automatic classification of it . Aone and Bennett (1995) report on work on Japanese, especially the treatment of zero-pronouns. Modejska et al (2003) report on experiments done on solving other-anaphora, that is, referential noun phrases modied by \other" or \another" and having non-structural antecedents. Finally, Soderland and Lehnert (1994a; 1994b) describe the implementation of a system for inter-sentential reference generation in the setting of an information extraction system.

2. Recasting the problem

The problem of resolving references is often perceived as identifying chains of referents, often across sentence boundaries, and sometimes even across documents. To make the problem more manageable for machine learning algorithms, it is often recasted as a classication and a clustering task. A classier determines whether or not two potential referents are coreferent or anaphoric, and a clustering mechanism coordinates the pairwise coreferent items into partitions, each of which contains the items that refer to the same entity. Since coreference is an equivalence relation, and anaphora is not, the two types of reference need somewhat different treatment. Coarsely, it can be argued that clustering is always performed implicitly in the case of anaphoric reference resolution, that is, since one is bound to look for an antecedent noun phrase rather than a pronoun that points to a noun phrase, what is obtained is a cluster of pronouns referring to the same noun phrase, rather than a chain of referents pointing back to a noun phrase. Thus anaphora resolution, as reported on by Aone and Bennett (1995), Connolly et al (1997), and Ge et al (1998), may be seen as a special case of coreference.

…

---

,

	Author	volume	Date Value	title	type	journal	titleUrl	doi	note	year
2004 ASurveyOfMLForRefResolutionInText	Fredrik Olsson			A Survey of Machine Learning for Reference Resolution in Textual Discourse			http://eprints.sics.se/2342/			2004