Focus Models

The focus model, or model of attentional state [10], is a model of which entities the dialog is most centrally about at each point in the dialog. It determines which previously mentioned entities are the candidate antecedents of anaphoric references. As such, it represents the role that the structure of the discourse plays in reference resolution.

We consider three models of attentional state in this paper: (1) the linear-recency model (see, for example, the work by Hobbs [14] and Walker² [39]), (2) Grosz and Sidner's [10] stack-based model, and (3) the graph structured stack model introduced by Rosé, Di Eugenio, Levin, and Van Ess-Dykema [31]. Ordered from (1) to (3), the models are successively more complex, accounting for increasingly more complex structures in the discourse.

In a linear-recency based model, entities mentioned in the discourse are stored on a focus list, ordered by recency. The corresponding structure in the dialog is shown in Figure 4a: a simple progression of references, uninterrupted by subdialogs.

  

Figure 4: Discourse Structures Targeted by Different Focus Models

In Grosz and Sidner's stack-based model, the entities in focus in a particular discourse segment are stored together in a focus space associated with that segment. To handle anaphoric references across discourse segments, focus spaces are pushed on and popped off the stack as appropriate to mirror the structure of the discourse. As each new segment is recognized, a focus space is created and pushed onto the stack. To interpret an anaphoric reference, the entities in the focus space on the top of the stack are considered first. However, if the current utterance resumes a previous discourse segment, the intervening focus spaces are popped off. This allows anaphoric reference to an earlier entity, even if more recently mentioned entities are possible antecedents (for more details, see [10]). Figure 4b illustrates a discourse structure that the stack-based model is designed to handle. Suppose that both TU₁ and TU₂ are possible antecedents of TU₃(for example, suppose they are specified by pronouns that agree in number and gender), but TU₂ is in a subsegment and is not a correct antecedent of TU₃, even though it is mentioned more recently than TU₁. In the stack-based model, the focus space containing TU₂ is popped off the stack when the end of its segment is recognized, thus removing TU₂ as a competitor for understanding TU₃. Following is an example from the NMSU corpus (this is the dialog segment labeled 09-09, in row 7, in Figure 10 presented later).

Dialog Date: Monday 10 May
	1	S1	Listen, daughter, I was thinking of inviting you to a demonstration on
			interior things, ornaments for decorating your house.
	2		Uh, I would like to do it at two p.m. Wednesday,
	3		But I don't know if you are free at that time or $\ldots$
TU1	4	S2	Uh, Wednesday, Mom, well
			Resolved to Wednesday, May 12
	5		You know that,
TU2,1	6		last week uh, I got a job and uh, a full-time job
			Unambiguous deictic; resolved to the week before the dialog date
TU2,2	7		I go in from seven in the morning to five in the afternoon
			Habitual
	8	S1	Oh, maybe it would be better
TU3	9	S2	Well, I have lunch from twelve to one
			Utterance (4) is needed for the correct interpretation:
			12-1, Wednesday 12 May

In this passage, utterances (6)-(7) are in a subdialog about S2's job. To interpret ``twelve to one'' in utterance (9) correctly, one must go back to utterance (4). Incorrect interpretations involving the temporal references in (6) and (7) are possible (using the co-reference relation with (6) and the modify relation with (7)), so those utterances must be skipped.

Rosé et al.'s graph structured stack is designed to handle the more complex structure depicted in Figure 4c. We will return to this structure later in Section 8.1, when the adequacy of our focus model is analyzed.

Once the candidate antecedents are determined, various criteria can be used to choose among them. Syntactic and semantic constraints are common.