Definite Description Processing in Unrestricted Text
Noun phrases with the definite article the, that we call DEFINITE DESCRIPTIONS, following (Russell, 1905), are one of the most common constructs in English, and have been extensively studied by linguists, philosophers, psychologists, and computational linguists. In this dissertation we present an implemented model of definite description processing that is based on extensive empirical studies of definite description use and whose performance can be quantitatively measured. In almost all approaches to discourse processing and discourse representation, definite descriptions have been regarded as anaphoric1; and the models of definite description processing proposed in the literature tend to emphasise the role of common-sense inference mechanisms. Recent work on discourse interpretation (Carletta, 1996; Carletta et al., 1997; Walker and Moore, 1997) has claimed that the judgements on which a theory is based should be shared by more than one subject. On the basis of previous linguistics and corpus linguistics work, we developed several annotation schemes and ran two experiments in which subjects were asked to annotate the uses of definite descriptions in newspaper articles. We compared their annotations and used them to develop our system and to evaluate its performance. Quantitative evaluation has become an issue in other language engineering tasks such as parsing, and has shown its usefulness also for theoretical developments. Recently, evaluation techniques have been introduced for semantic interpretation as well, as is the case for the Sixth Message Understanding Conference (MUC-6) (Sundheim, 1995). However, in this case, the emphasis was on the engineering aspects rather than on a careful study of the phenomena. Our goal has been to develop methods whose performance could be evaluated, but that were based on a careful study of linguistic evidence. The empirical studies we present are evidence that definite descriptions are not primarily anaphoric; they are often used to introduce a new entity in the discourse. Therefore, in the model of definite description processing that we propose, recognising discourse new descriptions plays a role as important as identifying the antecedent of those used anaphorically. Unlike most previous models, our system does not make use of specific hand coded knowledge or common-sense reasoning techniques; the only lexical source we use is WordNet (Miller et al., 1993). As a consequence, our system can process definite descriptions in any domain; a drawback is that our coverage is limited. Nevertheless, our studies serve to reveal the kind of knowledge that is needed for resolving definite descriptions, especially the bridging cases. The system resulting from this work can be useful in applications such as semi-automatic coreference annotation in unrestricted domains.