📄 Pragmatic Constraints on Semantic Presupposition

Summary

Work on presupposition in Discourse Representation Theory (DRT) is often criticised as purely semantic and truth-conditional. Critics such as Simons and Werth say it ignores what speakers and hearers believe and intend. Al-Raheb argues that this charge is partly unfair, but that DRT still needs a more agent-centred account. She redefines presupposition as a property of the speaker. On this view, a presupposition shows what the speaker believes or accepts, and what she thinks the hearer believes or accepts. It may even carry information that is new to the hearer, as long as it is presented as background. Grice's quantity and quality maxims are then recast as explicit belief constraints, separately for producing and for interpreting utterances. These constraints are implemented over an extended DRT representation (Al-Raheb 2005). In it, nested DRSs record each agent's beliefs and acceptances, along with that agent's model of the other agent. A short dialogue illustrates the approach. No quantitative evaluation is given. The contribution is a precise, checkable link between presupposition and agents' mental states.

Key points

• The paper separates sentence (semantic) presupposition from agent presupposition, which comes in a speaker variant and a hearer variant.
• It treats "given" broadly: a speaker may present new information as if known, to keep it out of focus. Unusual content still causes trouble (an ordinary car versus a small jet parked opposite the speaker's house). This view departs from Stalnaker's common-ground account.
• It distinguishes two attitude strengths:
  • BEL, which is belief;
  • ACCEPT, which is weak belief.
• There are two constraints on assertion:
  • BCA1 (quantity): the speaker must think the hearer does not already hold A. "Yeah I know" signals a breach.
  • BCA2 (quality): the speaker must herself believe or accept A.
• There are three constraints on presupposition, each with speaker-side and hearer-side checks:
  • BCP1: the presupposition must lie in the speaker's belief or acceptance space.
  • BCP1a: a hearer-side variant that updates the belief spaces.
  • BCP2: if the hearer is not yet known to hold the content, it is added to the hearer's acceptance DRS (accommodation).
• The implementation uses Prolog-style DRSs with nested attitude(i/you, 'BEL'/'ACCEPT', drs) conditions. It is demonstrated on a Vincent/Mia gift dialogue. If any constraint fails, the dialogue does not proceed.

Interesting ideas

• I1. Each agent keeps nested belief spaces: "what I believe" and, inside it, "what I think you believe", each split into firm belief and weaker acceptance. This makes it explicit whose mind a fact lives in.
• I2. Accommodated presuppositions go into the weaker acceptance space rather than into firm belief. This keeps a trace of how a fact entered the model.
• I3. The informativity check (BCA1) turns a reply like "Yeah I know" into evidence that the hearer already held the content. It is a concrete, belief-based reading of a Gricean maxim.
• I4. Unusual content packaged as background (the jet example) is expected to cause trouble, so the plausibility of presupposed material is a separate signal.
• I5. Each turn is gated by both a speaker-side and a hearer-side constraint check, and processing halts on violation. This gives a symmetric, inspectable account of production versus comprehension.

For Mova

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

• I1 (world): Text-world entities and states could carry a holder tag and a strength tag, so that a character's beliefs about another character are stored apart from narrator facts.
• I2 (world): Facts inferred from definite descriptions such as "Vincent's wife" could be added as accepted rather than asserted, so later contradictions can retract them cheaply and explainably.
• I3 (frames): The prag crate could flag responses that mark prior knowledge and update the speaker's model of the listener, which supports implicature detection in dialogues.
• I4 (frames): Presuppositions that clash with compiled world knowledge could be marked as remarkable, which is a cue for irony, surprise or a deliberate information trick.
• I5 (back): The generator could run the speaker-side checks before emitting a sentence, so it never presupposes content absent from its own world model.