The Use of Dynamically Extracted Context for Anaphoric Reference Resolution

1liE USE OF DYNAMICALLY EXTRACTED CONTEXT FOR ANAPOORIC REFERrnCE RESOLliTION

David Klappholz and Abe Lockman Department of Electrical Engineering and Computer Science Columbia University New York, New York 10027

We define anaphora more broadly than is usually done, i.e. to include not only pronoun, proadverb (e.g. here, there, then), pro-verb (e.g. do as in John goes to~early; I do too.), definite, etc. reference, but also indefinite reference (as in 1 below),as well as reference by an entire sentence or subsentence to a previous sentence or subsen­ tence (as in 2 below). In 1 an extremely inter­ esting tatooed lady is a reference to the circus, with the 1ntended relation ''was performing in" or something similar; In 2 My car wouldn't start is viewed as reference to I was late to work yester­ ~ with an intended relation which, to sweep the mrt under the rug' might be termed "cause"'' but which in fact is far more complicated. An ana­ phoric reference in this extended sense is thus an implied relation from an item (word, phrase, sub­ sentence, sentence) in a text to one or more pre­ viously occurring items (word, phrase, subsentence, sentence, group of sentences) in that text. Re­ solution of anaphora involves the identification of the latter item(s) as well as of the relation intended by the writer/speaker. Thus the resolu­ tion of all instances of anaphora in a text amounts to the determination of that part of the "meaning" of the text which is more than the collec­ tion of the "meanings" of its indiviudal sentences taken in isolation.

Since, in general, there is no test as to whether a particular item in a text is or is not an ana­ phoric reference (even a definite noun phrase need not be-~e.g. the first president of the United States In most contexts refers to a particular person, but not necessarily to a previously occur­ ring item in the text), an anaphora resolver must be a "guess and test" type algorithm. For each i~em in a ~ew sentence, a resolver must hypothe­ Size_r:lations between the item and some, hopeful~ ly limited, set of possible referents, and then use d~rected inferencing to judge which, if any, relation, to which, if any, possible referent was intended by the writer. This directed inferencing must use not only semantics and world knowledge, but also the structure of the previous text in order to properly resolve the reference.

w: are concerned with that aspect of an algorithi­ mic theory of anaphora resolution which is depen­ dent upon the "structure" of the text. Most pre­ vious approaches have either assumed that the structure of the text is known in advance or have ign?red it entirely. In particular we reject the notion that frames/scripts somehow "solve" the problem of anaphoric reference resolution.*

Most previous approaches make no real use of in­ ~ormation that the text itself provides concern­ lUg the flow of context. A successful algorithm must make use of the ways in which writers choose a particular form of presentation of their ideas in order to guide the flow of contexts and .allow the user to recover and use it in processing the text. These choices provide clues which must be used to dynamically extract local context at every point in the processing of a text. This local context is then a guide to reference resolution.

The most important factor in determining the local context is the way in which each sentence as a whole develops the material presented in the pre­ vious sentences of the text. Each sentence in a text is or contains either an expansion on (a further detail, adverbial, cause, effect, "how", etc.) or a temporal continuation of one or more previous sentences in the text (expansion can be either of a whole sentence or of one or more items within a sentence). In fact, the default assump­ tion made by readers is that a sentence expands or temporally continues the immediately previous sen­ tence of its text; semantics and world knowledge based inferencing must demonstrate that this link hypothesis is relatively weak in order to allow a sentence's interpretation as an expansion or tem­ poral continuation of a (several) sentence(s) further back in the text.

The set of expansion/temporal continuation re­ lations among the sentences of a text constitutes a graph which defines the context structure of the text. Our approach is to fit a new sentence into its proper places in the context graph repre­ senting the previous text, a fit that is accom­ plished by searching up the graph and hypothe­ sizing and testing sentence relations (expansion/ temporal continuation). Item to item reference hypotheses are entailed by the sentence relation hypotheses, and validated if and only if the re­ lations entailing them have been validated.

Finally we have demonstrated that, for a number of syntactic constructions, the choice that a writer makes as to how to group his semantic predicates into a surface sentence influences the way in which a reader will resolve later references to items in the sentence.

An integration of the above ideas into a sketch of an algorithm for resolving reference can be . found in Columbia University AI Project Technical Report 77- 2-1.

1. I went to the circus yesterday. I saw an ex­ tremely interesting tatooed lady.

2. I was late to work yesterday. My car wouldn't start.

* Our ideas can, however, be taken as the rudi­ ments of a theory of how frames/scripts can be dynamically shifted in and out of focus under control of a text's structure.