Coreference Resolution

As Redeker pointed out, the information obtained from a coherence analysis of texts is generally available by other means. One of the most common means is via coreference resolution. Let us first look at a coreference problem in local pragmatics, the resolution of a definite noun phrase (from IA, Section 5.1).

(4)	John walked into the room. The chandelier shone brightly.

What chandelier? Suppose we have axioms in our knowledge base specifying that rooms have lights in them:

(5)	$room(x) \, \supset \,light(y) \& in(y,x)$

and that a light with a branching fixture is a chandelier:

(6)	$light(y) \& branching$- $fixture(y) \, \supset \,chandelier(y)$

We know from the first sentence that there is a room. Axiom (5) allows us to infer a light in the room, and by assuming that that light has a branching fixture, we have the chandelier that we needed to find in the second sentence. What we had to assume, that the light in the room had a branching fixture, is the new information conveyed by the definite noun phrase ``the chandelier''. Figure illustrates this interpretation. Boxes are drawn around propositions that are assumed.

  

Figure: Coreference Resolution in Example (4).

Now let us return to example (3). The relevant explicit content of the text is

(7)	happy'(e1,A), promotion(e2,A)

That is, there is an eventuality e₁ that is Ann's being happy, and there is an eventuality e₂ that is Ann's promotion.

Coreference resolution is not a special process in the IA approach, but rather just an instance of pragmatic strengthening, of finding the best abductive proof of the explicit content of the sentence. It thus works as well for information conveyed by verbs and indefinite noun phrases as it does for definite noun phrases. In this example, part of the content of ``promotion'' is already implicit in the happiness, just as part of the content of ``the chandelier'' was already implicit in the room.

The axioms we use to illustrate this example are not necessarily the best way of expressing the facts, but their structure serves to emphasize the similarity of examples (3) and (4). First, we need to encode the fact that if someone is happy, then that is (defeasibly) caused by something that is ``better'' for that person (for instance, better than it has been or better than it might be):

(8)	$happy'(e_{1},x) \, \supset \,better(e_{2},x) \& cause(e_{2},e_{1})$

That is, if e₁ is x's being happy, then there is (defeasibly) an e₂ that is some sort of ``betterment'' for x and e₂ is the cause of the happiness e₁. In addition, a better job for someone is a promotion:

(9)	$better(e_{2},x) \& job(e_{2},x) \, \supset \,promotion(e_{2},x)$

That is, if e₂ is x's job and also some sort of ``betterment'' for x, then e₂ is a promotion for x.

In interpreting the second sentence of (3), we try to prove that there is a promotion. Using Axiom (9), we then try to find the existence of a better job. By using Axiom (8), we could establish that there is something better if we can find happiness, and the first sentence provides that. The minimal abductive proof is then obtained by assuming that the something better is a job. That is the new information conveyed by the assertion of a promotion. This interpretation is illustrated in Figure .

  

Figure: Example (3) as Coreference Resolution.

As a side effect of finding the coreferential link between ``promotion'' and ``happy'', we have inferred the causal relation that we wanted. No explicit search for or recognition of a coherence relation was necessary. It would seem, therefore, that at least in some cases, it is not necessary to look for a coherence relation between two segments. It happens anyway.