📄 Integrating Gricean and Attentional Constraints

Summary

The paper asks how speakers choose between a pronoun and a fuller noun phrase when they mention an entity again, and how a hearer resolves such mentions. Passonneau examined every discourse anaphoric NP (N=1,233) in ten narrations from Chafe's Pear stories. The data were already coded for coreference, utterance location and discourse segments, with segment boundaries validated by naive annotators. Each NP was judged on how informative it is relative to its context. The finding: re-mentions seldom say more than is needed to pick out the referent, whether they are pronouns or phrasal NPs. When they do say more, the extra detail usually falls at a segment boundary or at a shift of attention inside a segment.

On this basis she extends Dale & Reiter's distinguishing-description model into a single procedure that handles full phrasal NPs as well as pronouns, whether the task is producing them or interpreting them. She then combines it with centering. Centering narrows the candidate referents and can also license extra informativeness, for example to block a garden-path pronoun. The work matters because it replaces separate treatments of pronouns and full NPs with one account built on informativeness and attention.

Key points

• Corpus: 10 Pear stories narratives with 1,233 discourse anaphoric NPs, split into 563 phrasal NPs, 544 explicit pronouns and 126 zero pronominals.
• Discourse segments were marked by 7 untrained subjects per narrative. A boundary counted as validated when at least 4 subjects agreed, and the significance of this agreement was tested with Cochran's Q.
• Informativeness is classified four ways using the features ±adequate, ±efficient and ±increasing:
  • well-specified,
  • over-specified,
  • under-specified,
  • over-determined.
• Only 68 of 1,233 NPs (5.5%) were over-specified, and just 1 was over-determined. Of the 68 over-specified cases:
  • 30% fell at segment onsets,
  • 42% fell at intra-segmental attentional shifts,
  • so 72% coincided with some attentional shift overall.
• The link runs one way only: an over-specified NP tends to mark a shift, but a shift is usually not signalled by an over-specified NP.

Interesting ideas

• I1. The same distinguishing-description test checks a pronoun and a full NP, and it runs in both directions: choosing a form when generating and resolving a form when reading.
• I2. The context for a mention is widened step by step: first the entities of the previous utterance (Cf list), then the current focus space, then earlier or broader focus spaces, stopping at the first one that yields a unique solution.
• I3. An utterance with a null backward-looking center rules out definite pronouns, because a pronoun there would be read as the old center and cause a garden path.
• I4. Modifiers for a re-mention come from "focused attribute sets", for instance the attributes used in the most recent full NP for that entity.
• I5. Segment boundaries were taken from agreement among several naive annotators above a fixed threshold, and the reported correlation was explicitly checked in both directions.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): The coreference module could accept an antecedent only if the mention singles it out among the current candidates, using the same check for "he" and "the little boy".
• I2 (world): Entity lookup in the text world could search this nested sequence of scopes and record which scope succeeded, which makes each resolution explainable.
• I3 (back): The generator could choose a full NP whenever nothing from the previous sentence carries over, so the regenerated text does not mislead the reader.
• I4 (back): When the compressor turns the graph back into text, it could reuse the entity's last-mentioned attributes rather than choosing any attributes it knows.
• I5 (learn): Silver data from multiple witness parsers could use a similar vote threshold, and evaluations should test the reverse implication as a negative control.