Anaphora has to bind to a prior discourse object

What it is

A flight UA128 has just been confirmed; the user says “cancel it.” It has to land on the flight just introduced and still treated as present by both parties — not on the class “flights,” not on the airport mentioned last turn, not on the system’s default next departure. That process is anaphora binding to discourse referents. When binding fails the short word did not become vague — what is missing is an addressable entity for the prior object. Where candidates come from (ellipsis, a point on a screen, something in the room) is a different layer. This one is only: can “it / that / the one just now” be pinned to the thing already raised in the talk.

Why it happens

Each introducible object adds a discourse referent to common ground. Anaphora is a cheap pointer; people use it because they assume the pointer still works. A system that stores the previous utterance as a string, not an object identity, lets the pointer drift to the nearest noun or to the hottest slot in the intent classifier. Competing objects make the drift visible: UA128, then a change-fee policy, and “that” may be the flight or the policy. Humans resolve competition with salience, recency, and syntactic role; a nearest-noun rule will cancel the policy. Binding also requires the object to remain operable in state: a flight that has become a “cancelled” history item should bind to that history item, not to a freshly searched flight with the same number.

Studying it

Run anaphora resolution with controlled antecedents: one or two objects introduced over a couple of turns (a concrete flight, a concrete room), then “cancel it / move that one” at a controlled distance. Independents: antecedent distance (adjacent, or one intervening unrelated Q&A), presence of a competitor, whether the two objects are of the same type. Dependents: bind-to-correct rate, “which one” clarification rate, and the dangerous rate of binding to the competitor and acting.

From logs, pull user turns that contain a pronoun, label the discourse referent by hand, and see which slot the system actually changed. Do not stop at whether the intent was “cancel” — right intent, wrong object, is still a binding miss. Wizard-of-Oz can measure when people are willing to say only “it” before full entity linking exists.

Where it stops holding

If the current stretch has introduced no operable object, “it” has no antecedent; refuse to bind rather than guess a popular slot. With a single object the system itself just echoed (“UA128 is booked. change it?”), binding is effectively obligatory, and “cancel which flight” is false cleverness. If object identity is not shared across skills (calendar “it” versus mail “it”), binding cannot be done in engineering; say so, rather than bind to the other skill’s most recent item. “Remember one more sentence” is not enough: what must be remembered is the object, not the words.

Applying it

• Every successful act that later talk may point at leaves an identified object in state (which flight, which meeting, which order), not only an intent label.
• On a pronoun or “that,” resolve inside current common ground first. More than one candidate: ask a distinguishing property (time, destination); do not execute the nearest noun.
• Echo the bound identity in the confirmation (“cancel UA128”) so a wrong bind can be heard before action.
• How to check: a two-object competition script (book A, ask B’s policy, then “cancel it”). If the system touches B, or newly searches a same-number A, the bind did not pin.

Related

• Same group: M1.04.2 Context needs an explicit time-to-live · M1.04.3 Lost context must be announced, not silently reset
• Nearby: M1.09 Context Retention and Reference Resolution · M2.08 Explicit and Implicit Confirmation · M2.13 Multi-Intent and Compound Commands
• Search terms: anaphora binding to discourse referents · discourse referent · common ground