Utterance 1 Cf={John} (Cb unknown)
Utterance 2 Cb=John; Cf={John, Mike}
Utterance 3 Cb=John; Cf={John, Mike}
A variety of constraints have been proposed which appear to influence the resolution of pronouns and their felicitous usage. This section covers 3 approaches that reflect surface structure: Centering Theory, a Subjecthood bias, and Parallelism. Even though they all focus on surface properties (word order, syntactic position) that do not require deep inferencing about the events or situations being described or the larger discourse context, even structural constraints can be said to reflect deeper pragmatic properties. Recall from the section on Information Structure that syntactic options often reflect underlying pragmatic distinctions. Centering Theory is based around a surface metric of the "center" of a discourse, but it is intended to capture something deeper---namely that understanding a discourse requires tracking who or what is the topic of the discourse from utterance to utterance. Subjecthood and Parallelism biases reflect researchers' observations that subject position is a typical position for topical information to appear and that parallel structure favors parallel interpretation.
Centering Theory is a model that comes out of computational linguistic traditions and was originally a model for language generation (Grosz, Joshi, & Weinstein, 1995). It centers around the claim that the coherence of a discourse depends on inference demands made by different types of referring expressions. Examples (7) and (8) show two different versions of an excerpt that we will assume is part of a longer segment centered on the referent John. The difference between (7) and (8) is the referring expression used to refer to John in the last sentence.
(7) HeJohn has been acting quite odd.
HeJohn called up Mike yesterday.
John wanted to meet himMike urgently.
(8) HeJohn has been acting quite odd.
HeJohn called up Mike yesterday.
HeJohn wanted to meet himMike urgently.
The authors of Centering Theory have noted that the use of the proper name John in the last line of (7) makes (7) harder to understand. Successful interpretation of (7) almost seems to require a 2nd John other than the one under discussion, whereas 8) easily accommodates a single John. It is this intuition --- that there are more and less appropriate forms to use when re-mentioning particular referents --- that underlies Centering Theory. Thus, referring expressions can serve to inform the hearer: "the use of a pronoun ... signals [to] the hearer that the speaker is continuing to talk about the same thing" (Grosz et al. 1995). This informativeness and resulting coherence of the discourse is achieved via a set of rules and constraints: "Certain entities in an utterance are more central than others and this property imposes constraints on a speaker’s use of different types of referring expressions... The coherence of a discourse is affected by the compatibility between centering properties of an utterance and choice of referring expression."
Centering Theory sets out the definition of two important components of a discourse: the list of Forward-looking centers Cf for a particular utterance Un and a single element that is the Backward-looking center Cb between a pair of utterances.
Forward-looking center list Cf(Un): a partially ordered list of entities mentioned in utterance Uu such that subject > object(s) > other (where ">" means "outranks")
Backward-looking center Cb(Un+1): the most highly ranked element of Cf(Un) that is realized in Un+1
Centering Theory also posits a rule regarding pronominalisation:
Rule 1: If any element of Cf(Un) is realized by a pronoun in Un+1, then the Cb(Un+1) must be realized by a pronoun also.
For examples (7) and (8) above, repeated below as (9), we can identify the forward-looking center list (Cf(Un)) for the 1st utterance. It consists of the list of entities mentioned; in this case there's just one entity, John: Cf = {John}. For the 2nd utterance, we can identify both the backward-looking center and the forward-looking center list. The backward-looking center must be an element that is both listed on the previous utterance's forward-looking center list and mentioned in the current utterance. For the 2nd utterance, the Cb=John.
(9) HeJohn has been acting quite odd.
HeJohn called up Mike yesterday.
John / HeJohn wanted to meet himMike urgently.
Centering Theory has several more definitions and one more rule. The following are the definitions for the three types of transitions between utterances, defined in terms of the maintenance/change in topic.
Center continuation: Cb(Un+1) = Cb(Un) and this entity is the most highly ranked element of Cf(Un+1)
Center retaining: Cb(Un+1) = Cb(Un) and this entity is not the most highly ranked element of Cf(Un+1)
Center shifting: Cb(Un+1) ≠ Cb(Un)
Rule 2 is a rule that ranks the above transition types in order of coherence and ease of processing.
Rule 2: continuation > retaining > shifting
To see these definitions in action, consider the following passage in (10).
(10) John has been having a lot of trouble arranging his vacation.
He cannot find anyone to take over his responsibilities.
He called up Mike yesterday to work out a plan. [continuation]
Mike has annoyed him a lot recently. [retaining]
He called John at 5am on Friday last week. [shifting]
Note that the reason that the theory is spelled out in these formal terms is in part because it was intended as an algorithm that could be implemented for automated reference resolution. Structural constraints are much easier to include in a computational system since they don't require world knowledge or pragmatic inferencing. There are two other structural biases that have been observed (primarily in the psycholinguistic literature).
One is the Subjecthood constraints which has been posited as a way to account for an observed bias towards antecedents that have appeared syntactically in subject position (Kamayama, 1986). This constraint was introduced based on the existence of data like the pairs of sentences in (11) and (12). A natural interpretation of the pronoun "him" in (11) is in reference to John. When the same event from (11) is described with the passive in (12), the preferred referent of "him" seems to switch from John to Bill. It was data like this (in which the underlying semantic content of the kicking event remains the same but the structural position of the referents changes) that led to the suggestion that maybe pronouns prefer to find their antecedent in subject position.
(11) John kicked Bill. Mary told him to go home.
(12) Bill was kicked by John. Mary told him to go home.
However, in contrast to the subjecthood constraint, there is also data that points to what has been called a Parallelism constraint, namely a constraint to prefer antecedents in syntactically parallel positions (Smyth 1994). In (13) and (14), the pronoun seems to favor the parallel referent--the object in (13) and the subject in (14).
(13) John kicked Bill in the leg. Mary punched him in the arm.
(14) John kicked Bill in the leg. He punched Mary in the arm.
Again, both Centering Theory as well as the Subjecthood and Parallelism constraints function by tapping into surface properties of the passages in which pronouns are found. They do not require deeper inferencing. As such, they are appealing from a computational modeling, simple heuristics approach, but unfortunately there is a set of interesting data that they fail to capture. The next section discusses contexts that show that other biases that are at play which depend on meaning rather than surface structure.
To go on to section 8.3 "Semantic factors in reference", click here.