📄 Against the Identification of Anaphora and Presupposition

Summary

This position paper questions the near-consensus, built on work by van der Sandt and Geurts in Discourse Representation Theory, that presupposition and anaphora are basically the same phenomenon. Bosch calls this the PIA ("Presupposition is Anaphora") hypothesis. He agrees the two are related but argues that merging them is empirically false and blurs both concepts.

He offers four groups of data that the hypothesis handles poorly:

• definite NPs whose uniqueness comes from their own lexical and compositional meaning, with no discourse antecedent;
• presuppositions built into predicates and discourse particles;
• epithet NPs that pick up a known referent while their literal content is ignored;
• German pronouns whose grammatical gender is meaningless unless the pronoun is contrastively stressed.

His alternative picture separates the two. Anaphora maintains reference across a discourse. Presupposition is an inferential mechanism that limits which contexts are acceptable. The paper ends by suggesting a key difference: presupposition may be modelled compositionally, while anaphora likely cannot be. This is an argument for keeping them as separate mechanisms in theory and in processing.

Key points

• The paper takes aim at the PIA hypothesis as set out in Geurts (1995, 1999), in van der Sandt's 1992 work, and in the joint 1991 paper by van der Sandt and Geurts, which DRT implements through binding and accommodation.
• Definite NPs follow Löbner (1985): a definite denotes a function with a unique value. Context-driven cases are resumption, bridging and utterance-situation reference. Purely internal cases include "the sum of 3 and 4" and "the H2O molecule".
• Lexical presuppositions with no plausible anaphora:
  • bachelor → adult
  • ignore → notice
  • close the door at 5 → door open at 5
  • particles such as but, although, even, anyway
• Epithets such as "the old grocer" behave in two incompatible ways:
  • Unstressed, the NP refers back to Jones and only conveys the speaker's attitude, with no effect on truth conditions.
  • Focused (stressed or clefted), the same NP introduces a new referent who really is an old shopkeeper.
• In German (Mutter/Bolzen vs. Nachbar/Freundin), unstressed pronoun gender is purely formal agreement. Under contrastive accent, gender is read as sex, which makes stressed reference to inanimate objects unacceptable.

Interesting ideas

• I1. Some definites are unique by their own description, so forcing an antecedent search and "accommodating" them erases the line between lexical and discourse semantics
• I2. An NP has two mutually exclusive readings: as an anaphor its content is idle, and as a presupposition trigger its content becomes a real property of a new referent
• I3. Focus marking, such as an it-cleft "it was X who", blocks the anaphoric reading and pushes toward introducing a new referent
• I4. Lexical presuppositions of predicates are separate from reference: "close" presupposes "was open"
• I5. Presupposition may compose bottom-up, while anaphora resists compositional treatment

For Mova

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

• I1 (world): Definites like "the sum of 3 and 4" should create a new entity directly instead of failing a coreference lookup. The trace should record why no antecedent was sought.
• I2 (frames): The coreference module should flag epithets ("the fool", "the poor fellow") as attitude-only. Their lexical content should then not be asserted as properties of the linked entity.
• I3 (frames): In plain text, cleft constructions can stand in for missing prosody. They can serve as a cue against resolving an epithet to the salient antecedent.
• I4 (world): Store presupposed prior states per verb class in the compiled knowledge. Event order can then infer the earlier state (door open before 5) as an explainable inference, not a coreference link.
• I5 (frames): Compute presuppositions compositionally on the UD tree. Treat coreference as a separate discourse-level process, so each step's justification stays clear.