Contribution
The paper gives Discourse Representation Theory’s merge-and-unify step a sheaf-theoretic form. Basic DRS become sections of a presheaf F on a category of (vocabulary, variables) pairs; DRT’s identification of referents becomes a cover; and the semantics of a whole discourse is the gluing of its sentences’ sections along that cover. It then composes F with a distribution functor to accommodate preferential anaphora, where several resolutions are possible. It is a programme paper: one proposition, worked examples, and three directions for future work (the logical operations of full DRS, complexity of finding a gluing, learned distributions).
Key insight
Resolving anaphora is choosing how to identify the referents of separate sentences, and once that identification (the cover) is fixed there is nothing left to choose: the global meaning either exists, uniquely, or the identification is semantically wrong. Interpretation lives in the cover, and the gluing condition is a correctness check on it.
Assumptions
- A first-order language with relation symbols only: no constants, no function symbols.
- Basic DRS only: conditions are finite sets of literals. Negated, conditional, disjunctive and quantified conditions are left to future work, so accessibility (the donkey sentence’s real difficulty) is discussed in §3 but never modelled.
- Sections are deductive closures of consistent finite sets of literals; inconsistency is what takes a candidate out of F.
- No Grothendieck topology is defined. Covers are given concretely as jointly surjective families {fᵢ : (Lᵢ, Xᵢ) → (L, X)} with L = ∪ Lᵢ, and gluing is defined directly by restriction from the global section. There is no separate compatibility-on-overlaps condition, so “compatible family” has no independent meaning here.
- In every linguistic example the local vocabularies are disjoint, which by the paper’s own remark makes consistency the only obstruction.
Key results
- Presheaf F (§4.1). F(L, X) = deductive closures of consistent finite sets of literals over X in L; for f : (L, X) → (L′, Y), F(f)(s) ⊢ ±A(x) iff s ⊢ ±A(f(x)). Functoriality is asserted (“easily verified”).
- Proposition 1 (§4.2). For a cover {fᵢ} and family {sᵢ}, a gluing, if it exists, is unique: it can only be the deductive closure of {±A(fᵢ(x)) | ±A(x) ∈ sᵢ}, and it is a gluing if consistent and if it restricts to each sᵢ. Holds when: any cover; existence is not guaranteed.
- Failure without inconsistency (§4.2). With L₁ = L₂ = {R, S}, s₁ = {R(x), S(u)}, s₂ = {S(y), R(v)}, x, y ↦ z and u, v ↦ w, the candidate restricts along f₁ to {R(x), S(x), R(u), S(u)} ≠ s₁. With shared vocabulary, a gluing fails if the merge adds anything about a local referent in its own vocabulary. I checked this; it is right.
- Examples (§4.3): “John sleeps. He snores.” (one cover, glues); “John beats his donkey” (three local sections glued onto {a, b}); “John owns a donkey. It is grey.” (merging “it” with John is inconsistent via Man/¬Man, merging it with the donkey glues); “John put the cup on the plate. He broke it.” (two covers, both glue).
- Distribution functor (§5). D_R over a commutative semiring R: finite support functions summing to 1, with image measure as the action. Over the Booleans it is the finite powerset functor, over R⁺ finite probability distributions. Same pattern as LIT-016’s D_R ℰ, with F in place of the event sheaf ℰ.
- Ranking covers (§5.1). For “John gave the bananas to the monkeys. They were ripe. They were cheeky.”, the four covers c₁–c₄ get probability proportional to the sum of the corpus counts of their two adjective–noun patterns (ripe banana 14, ripe monkey 0, cheeky banana 0, cheeky monkey 10): 14/48, 24/48, 0, 10/48. The printed d(t₄) = 0.205 is 10/48 = 0.208. c₂ (bananas ripe, monkeys cheeky) is chosen.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Basic DRS form a presheaf on (vocabulary, variables) with restriction by pulling literals back along variable maps | moderate | §4.1; functoriality asserted, not shown |
| C2 | Relative to a fixed cover, semantic unification has at most one result | strong (proof) | Proposition 1 (immediate from the definitions) |
| C3 | DRT’s merging of referents is the choice of cover, and the gluing condition expresses the semantic correctness of the merge | moderate | §4.3 examples; an interpretation, illustrated, not a theorem |
| C4 | Constraint-based anaphora resolution (agreement via negative literals) is the existence or non-existence of a gluing | moderate | Example 3 of §4.3 |
| C5 | Composing with a distribution functor handles preferential anaphora | weak | §5.1: the distribution is placed on the candidate global sections directly, not obtained by gluing local distributions; maximum entropy is mentioned and not used |
| C6 | The corpus-event probabilities “correspond to the coverings of the sheaf model” | weak | one worked example; the additive scoring gives 14/48 to “the bananas were cheeky” although “cheeky banana” has count 0 |
| C7 | The approach extends to full DRS and “to some extent obviates the need for” its extended conditions | not supported here | stated as belief in §3 and as future work in §6 |
Concepts
- section: an element of F(L, X), a deductively closed consistent set of literals; a basic DRS.
- cover: a jointly surjective family of maps into (L, X), with the vocabularies’ union L. Covers can identify variables, so a cover is a resolution of the anaphors.
- gluing / semantic unification: a section over (L, X) restricting to every local section along the cover.
- contextual: used throughout in the sense of context-dependent meaning (Harris and Firth’s distributional context, and the discourse context of an anaphor), not in the sense of a compatible family with no global section.
- multivalued gluing: the case where several gluings, or several covers’ gluings, are possible and must be ranked.
Connections
The sheaf machinery and the D_R functor are from Abramsky and Brandenburger (LIT-016) and Abramsky’s work on relational databases; the linguistic side is DRT (Kamp and Reyle) with DPL and Visser’s monoid semantics named as compositional successors. The donkey sentence comes from Geach. The probabilities follow Dagan and Itai (COLING 1990). The paper mentions DisCoCat (LIT-273) only in the introduction, as another way language is “contextual”; there is no construction linking the two, and no use of vector-space meanings.
Bearing on the record
- LIT-016 / NOTE-016, THEORY-012. The paper borrows the presheaf, gluing and distribution-functor shape of LIT-016, but not its notion of contextuality. In LIT-016 and THEORY-012, contextuality is a family that agrees on overlaps and still has no global section. Here no overlap-compatibility is defined, and the obstructions shown are inconsistency (Example 3) and over-determination of a local section (the R, S example). Neither is that phenomenon. The paper is the source of the “sheaves for language” idea; the record should not cite it as showing that language is contextual in THEORY-012’s sense.
- LIT-277, LIT-265. Nothing here computes a cohomological obstruction or a contextual fraction, and neither could apply as things stand: both measure the failure of a compatible family to glue, and this paper’s families have no overlap condition to be compatible under.
- THEORY-013. That document separates context-dependent marginals from contextuality in language data. This paper is an early instance of the conflation it warns about: its abstract says language is contextual and sheaves model contextuality, but what it models is context dependence and ambiguity. The paper has no data, so it neither supports nor contradicts THEORY-013’s empirical claim.
- LIT-273 / NOTE-245, LIT-272 / NOTE-246. These are the same co-author’s compositional distributional programme. The two approaches are presented side by side and not joined. NOTE-245 finds that DisCoCat’s posetal framing cannot distinguish parse ambiguities. Here anaphoric ambiguity is represented by a set of covers, so the two programmes place ambiguity in different structures, and neither paper relates them.
- LIT-838 / NOTE-608, THEORY-162. Heim’s file change semantics is DRT’s sister account of indefinites and donkey anaphora. This paper’s cover does the work Heim gives to a pronoun finding a familiar card. It handles only the basic, unembedded case, so the accessibility facts that motivate both DRT and Heim lie outside it.
- No THEORY filed. The reading’s one structural point is that ambiguity sits in the choice of cover and a fixed cover glues at most once. That is a fact about this construction, stated as Proposition 1, not a finding about language that the record lacks.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Only basic DRS; no negation, implication, disjunction or quantifier conditions, so none of the accessibility constraints that DRT exists to handle. The authors say so (“rather short and preliminary”).
- No Grothendieck topology, so F is not shown to be a sheaf, and the cover notion is ad hoc.
- How covers are generated (which identifications are candidates) is outside the model; the paper calls it the “intelligence” of the operation and leaves it to accessibility constraints and corpus counts.
- The probabilistic section does not glue local distributions. It places a distribution on candidate global sections by an additive score, without normalising per anaphor or assuming independence. A product of counts would give c₂ probability 1, and the additive score gives weight to readings with a zero-count pattern.
- One worked corpus example; no evaluation. Small errors: Example 2 is labelled “intersentential” although anaphor and antecedent are in one sentence, and 10/48 is printed as 0.205.
Open questions
- Whether a genuinely contextual discourse exists in this framework: a family of local DRS sections that agree on every overlap of a fixed cover and have no gluing. That would need shared vocabulary and an overlap condition the paper does not define.
- Whether the cover-choice step can be internalised, for example as a presheaf over covers or a sheaf on a site whose objects include identifications, so that ranking resolutions becomes gluing rather than an external score.
- The extension to full DRS that §6 promises. Later work by Sadrzadeh and collaborators on contextuality in ambiguous phrases may take this up; it was not checked for this note.