Early Approaches to Anaphora Resolution: Theoretically Inspired and Heuristic-Based

This chapter summarizes the most influential non-statistical approaches to anaphora resolution. Much of the very early work focused on personal pronouns and was based on theoretical proposals concerning anaphora and its interpretation developed in linguistics (e.g., the effect of syntax or semantics on anaphora) and/or psychology (e.g., on the effect of salience or commonsense knowledge). Such systems assumed the resolver would have perfect information available – e.g., on the syntactic structure of the sentence, or the properties of concepts and instances – and as a result, tended to be very brittle (a notable exception being Hobbs’ ‘naive’ algorithm for pronoun resolution). In the first part of this Chapter we cover in detail some of these theoretically-motivated algorithms, such as Hobbs’ and Sidner’s, and briefly survey a number of other ones. The availability of the first corpora in the mid-90s (see Chapter 4) led to the development of the first systems able to operate on a larger scale, and to a widening of the range of anaphoric expressions handled.
The fundamental property of these systems was the ability to carry out resolution on the basis of imperfect information only, using a variety of heuristics. In the second part of this Chapter, we cover a number of these heuristic-based algorithms. Some of the ideas developed in these heuristic-based systems have come back and are the basis for systems developed in the last few years; of these, we will discuss in some detail the Stanford Deterministic Coreference System.