nextupprevious
Next:The anaphora resolution algorithmUp:Anaphora resolution in SpanishPrevious:Anaphora resolution in Spanish


Constraints and preferences as an approach to anaphora resolution

According to Dahlbäck 1991, there are at present two basic approaches in anaphora resolution: (1) the traditional approach, which generally depends upon linguistic knowledge, and (2) the discourse-oriented approach, in which the researcher tries to model complex discourse structures and then uses these structures to resolve anaphora.

Among the traditional approaches, the work of Mitkov 1998, Baldwin 1997, and Ferrández et al. 2000 are all based on a combination of linguistic knowledge (lexical, morphological, syntactic, and/or semantic) for the resolution of anaphora. These approaches apply linguistic knowledge, in the way of constraints and preferences, following the work of Carbonell and Brown 1988 and Rich and LuperFoy 1988, in which such systems are proposed as a technique for combining several information sources.

These approaches are based, intuitively, on the following three steps: (1) defining an anaphoric accessibility space, (2) applying constraints, and (3) applying preferences.

A constraint and preference system must define, on the one hand, the anaphoric accessibility space. That is, it must obtain a list with all the possible candidate antecedents. On the other hand, the system must also define the text segments in which the antecedent can be found. This step has a great importance for the remaining steps in the process because a definition of the anaphoric accessibility space that is too narrow results in the exclusion of valid antecedents. Likewise, a definition of the anaphoric accessibility space that is too broad results in large candidate lists, with a corresponding increase in the likelihood of erroneous anaphora resolution. Usually, anaphora resolution systems based on linguistic knowledge [Ferrández et al.1999] define an accessibility space using n previous sentences to the anaphor, where n is variable according to the kind of the anaphora.

Once the list of possible candidates is defined, several constraints are applied in order to remove incompatible antecedents. The constraint system consists of conditions that must be met, and candidates that do not fulfill these conditions will not be considered possible antecedents for the anaphor. Lexical, morphological, syntactical, and semantic information are traditionally used to define the constraints.

Finally, after removing incompatible candidates, if the remaining list contains more than one antecedent, preferences are applied in order to choose a single antecedent. In this case, unlike that of constraints, preferences are associated with likelihood lower than 100%. Candidates fulfilling a preference, then, have a greater likelihood of being the antecedent than those not fulfilling it. The preference system must be designed bearing in mind that only one candidate must remain at the end. This final candidate will be proposed as the antecedent for the anaphor. Lexical, morphological, syntactic, and semantic information are usually used in order to define the preference system.

The works of Mitkov 1998 Mitkov98 and Ferrández et al. 2000 show that anaphora resolution systems based on constraints and preferences can yield successful results when applied to non-dialogue texts. However, these works lack adequate proposals for the anaphoric accessibility space. Furthermore, these approaches lack consistency in the treatment of other kinds of texts, for example, dialogues.


nextupprevious
Next:The anaphora resolution algorithmUp:Anaphora resolution in SpanishPrevious:Anaphora resolution in Spanish
patricio 2001-10-17