📄 An Integrated Approach to Reference and Presupposition Resolution

Summary

The paper argues that resolving pronouns and definite descriptions should not be a standalone task. Instead, both are special cases of a single process: satisfying presuppositions against a structured discourse context. Building on Roberts' framework of discourse as cooperative inquiry, the authors model context as an Information Structure. It holds the moves so far, the common ground (CG), a Domain of familiar discourse referents, and a stack of questions under discussion (QUD). The QUD stack mirrors the intentional hierarchy of a dialogue and narrows the search for antecedents and other presupposed material. The authors give algorithms for updating this structure and retrieving presupposed content, inside a partially implemented hotel-booking dialogue system. Four sample dialogues cover long-distance pronoun reference, a definite resolved via an indefinite in an earlier question, contextual domain restriction of "every," and an elliptical "How about X?" question. There is no quantitative evaluation. The contribution is a unified, explainable treatment of several context-dependent phenomena under one strategy.

Key points

• Definites (pronouns, demonstratives, "the"-NPs) presuppose two things. Weak familiarity means the CG entails that a matching entity exists, with no explicit antecedent required. Informational uniqueness means only one familiar referent fits the description, which is weaker than Russell's semantic uniqueness.
• Pronouns add a conventional requirement of maximal salience, implemented through the QUD stack. This explains why pronouns and definite descriptions are distributed differently.
• Relevance is defined formally. An assertion must give at least a partial answer to the top question. A sub-question's full answer must contextually entail a partial answer to the question beneath it.
• Several algorithms are presented, including resolve_WH_Ellipsis, a match-and-substitute step, add_domain_restriction, and a priority union of discourse representations. Logical forms are processed inside-out, so embedded presupposition problems are solved first.
• The authors list open problems:
  • coercing requests and need/desire statements into questions;
  • inferring the user's domain goals;
  • running a corpus evaluation, which is still pending and hard because of domain knowledge.

Interesting ideas

• I1. Treat every context-dependent construction as a presupposition to satisfy, with one shared search procedure; this replaces a separate module per phenomenon.
• I2. Keep a stack of open questions where only the top is writable but every entry is readable, and pop a question when it is answered or shown unanswerable. This gives a cheap, explicit model of topic hierarchy for long-distance reference.
• I3. Familiarity is licensed by what the common ground entails, not by a literal NP antecedent, so bridged or implied entities can still be referred to definitely.
• I4. Use descriptions already present in the open questions as an approximate domain restriction, instead of running costly full plan inference.
• I5. Resolve "How about X?" ellipsis by substituting X into the matching typed slot of the top open question; if the wh-slot is filled by a definite, the result becomes a yes/no question.

For Mova

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

• I1 (frames): coreference, ellipsis and quantifier domain restriction in prag could share one "find it in context" routine, which makes each resolution step traceable to the same rule.
• I2 (world): the world of a text could track open questions or goals alongside entities, so a pronoun many sentences later can be tied back to the still-open super-topic.
• I3 (world): entities inferred from states and knowledge (e.g. a hotel implied by a booking) should be registered as referable, letting "the hotel" resolve without a prior mention.
• I4 (answer): when a test question says "every morning" or "the box", the reader can restrict the domain using the content of the current question or task statement.
• I5 (frames): fragment answers and follow-ups in QA on fables could be rebuilt by slot substitution into the prior question's frame, with a fallback that pops to older questions when no slot fits.