{"id":"e40090b3-5fc1-4884-9853-d47e66147e4a","arxiv_id":"1908.01355","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Adding indexed contexts and scoping constraints to AMR yields AMR+, a formalism that resolves the bound variable and scope representation problems and translates to DRT-style representations.","lead":"This paper proposes AMR+, a version of Abstract Meaning Representation that adds context indices and scope constraints to handle negation, quantifiers, and bound variables. It offers a path for AMR to support logical inference by translating the new structure into Discourse Representation Theory.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DRS translation τ in §3.5 is under-specified and internally inconsistent with the scoping intended in §3.4, so the claimed correctness of AMR+ is not yet established.","rationale":"The reader's conditional verdict is well-calibrated. The most load-bearing assumption is indeed the faithfulness of the τ translation and the associated semantics for context constraints. The paper explicitly says details were left out, and the observed dog-index inconsistency between §3.4 and §3.5 confirms that the translation is not currently precise enough to support the central correctness claim. However, this is a fixable formalization gap rather than a demonstration that the approach is wrong: the examples are small, the intended readings are reasonably clear, and a more careful recursive definition of τ could likely resolve the issues. Thus the reader's CONDITIONAL verdict remains appropriate. I do not see a stronger objection that would require rejection, and the paper's contribution is modest but real: it identifies a genuine scope/bound-variable problem in AMR and proposes a concrete representational extension with close ties to DRT. The missing piece is formal rigor, not fundamental implausibility.","tokens_in":6151,"tokens_out":4235,"duration_ms":44046,"concrete_test":"Formalize AMR+ syntax and give a denotational semantics for constraints, then implement the three-step translation τ as a recursive function. Run all examples from §3.3–§3.4 through this implementation. In particular, translate both versions of 'A dog scared every cat': the §3.4 version with x/3 and constraint {3:2=>1}, and the §3.5 version with x/1 and the same constraint. Check whether the two translations are logically equivalent, and compare their truth conditions against the intended wide-scope reading. If the §3.5 translation places the existential dog inside the universal consequent, the constraint semantics and τ do not yet determine the reading claimed in §3.4.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that AMR+ 'makes correct predictions' for inferences involving negation and bound variables and is equivalent to PMB-style DRS. This requires the labelled-context notation to have a determinate semantics and the translation τ to be faithful. Neither condition is demonstrated. Constraint expressions such as {3:2=>1} are used ambiguously: the leading '3' is sometimes an inferred context (as in §3.3 and §3.4) and sometimes a label already attached to a sub-AMR (the dog x/3 in §3.4). No recursive semantics for 'l : C' is given, so it is undefined whether {3:2=>1} means 'the context containing the dog contains the implication', 'the whole formula consists of the dog plus the implication', or something else. The internal inconsistency between §3.4 (dog indexed x/3, intended wide scope) and §3.5 (dog indexed x/1) makes the problem concrete: applying τ to the §3.5 version puts the existential dog inside the consequent of the universal, yielding 'every cat was scared by some dog', not 'a dog scared every cat'. The claimed equivalence with PMB DRS therefore rests on an unverified mapping, and the correctness of the formalism for scope inferences is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AMR+, an extension of Abstract Meaning Representation that adds a logical layer on top of the usual predicate-argument structure. Each sub-AMR receives an index, and additional scoping constraints over these indices (=, ¬, ⇒, <) are used to represent negation, quantification, and presuppositions. The paper gives examples for existential and universal quantification, negation, proper names, and bound variables, and sketches a translation τ from AMR+ to the DRS dialect used in the Parallel Meaning Bank. The central claim is that AMR+ 'makes correct predictions' for inferences involving negation and bound variables and is equivalent in expressive power to that DRS dialect, while preserving AMR's attractive core structure. The paper is a short proposal: it does not give a model-theoretic semantics for the labelled contexts, and the translation τ is presented only by examples.","tokens_in":6491,"tokens_out":5406,"duration_ms":54600,"significance":"If made rigorous, this proposal would address two well-known limitations of AMR: the bound variable problem and the scope representation problem. The idea of separating predicate-argument structure from logical structure is simple and potentially useful, and the link to DRT could enable cross-formalism comparison and inference. The paper is honest about its open ends and does not overstate its results into a full formal theory. However, the paper's headline claims of correctness and DRS equivalence are not yet supported: the semantics of the constraint language is left informal, and the translation τ is under-specified and, as the examples show, internally inconsistent in a load-bearing case. The contribution is therefore best seen as a promising research program rather than a verified formalism.","major_comments":[{"comment":"The same sentence \"A dog scared every cat\" is represented with different variable indices in the two sections: in §3.4 the dog is x/3 and the constraint is {3:2=>1}, apparently giving the dog wide scope, whereas in §3.5 the dog is x/1, already merged into τ(1), and the translation yields τ(3)=τ(2)⇒τ(1), which places the existential dog inside the consequent of the universal. The resulting DRS reads as \"every cat was scared by some dog,\" not as \"a dog scared every cat.\" This inconsistency shows that the mapping from AMR+ to DRS is not deterministic and undermines the claimed correctness of the formalism for scope inferences.","section":"§3.4 and §3.5"},{"comment":"The translation τ is not a well-defined function. No rule is given for interpreting the leading label in an expression such as {3:2=>1}: the leading 3 is sometimes an inferred context (as in §3.3) and sometimes the label of a sub-AMR (the dog in §3.4). Similarly, no semantics is provided for the presupposition constraint <; in the \"Mary didn't smile\" example, τ(3)=¬τ(1) leaves the variable x in Agent(e,x) unbound, and the example does not show how τ(2) combines with τ(3). Without a recursive definition of l:C and a model-theoretic account of the constraints, the equivalence claim with PMB-style DRS is not testable.","section":"§3.5, translation τ"},{"comment":"The constraint set {2=3,3<1,4:2=>1} is unclear. In the AMR, x/2 is the student and y/3 is the paper, so 2=3 appears to identify the student's context with the paper's context, which is not the intended DRS reading. No explanation is provided of how these constraints encode the universal over students and the coindexation of the possessive pronoun with the student. This example suggests that the constraint language needs a precise semantics before the informal readings of the examples can be checked.","section":"§3.4, \"Every student revised their paper\""}],"minor_comments":[{"comment":"The phrase \"makes correct prediction for inferences\" is grammatically incorrect; it should be \"makes correct predictions for inferences.\"","section":"Abstract"},{"comment":"The description of the two-place relation says “the second argument of the two-place relation is the main variable of the sub-AMR” twice; the second occurrence should refer to the main variable of the argument sub-AMR. The displayed formula also uses 'x' for the main variable while the surrounding text uses 'i'.","section":"§3.5, Step 1"},{"comment":"The proposed shorthand {¬j} for {i:¬j} and {j⇒k} for {i:j⇒k} is introduced without a rule for choosing the containing context i; this is especially problematic when multiple inferred contexts are present, as in §3.4.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's own acknowledgements say that a formal definition of the translation was requested by a reviewer and that details had to be left out. Given that the examples themselves are inconsistent (the 'dog scared every cat' case), the correctness claim is not yet established. I would encourage the editor to ask for a proper formal definition of the context semantics and a full worked-out translation, or at least a careful revision of the examples. The paper is a proposal and the core idea is worth developing, but in its current form it is not ready for publication as a rigorous claim of equivalence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nYou can skip the proofs and read the examples; this is a short position paper about adding a logical layer to AMR. The central idea is plausible: keep the predicate-argument structure, add indices to sub-AMRs, and add constraints like 2:¬1 and 3:2=>1 to capture negation and quantifier scope. For simple sentences the notation works nicely, and the paper gives a clear account of why plain AMR fails on bound variables and scope.\n\nThe genuinely new bit is the particular constraint notation and the claim that AMR+ can be translated into the DRS dialect used in the Parallel Meaning Bank. If that translation were solid, it would let AMR inherit DRT's inference machinery. But it is not solid as written. The translation τ is sketched in three bullets, not defined recursively; the paper explicitly says details were left out. More concretely, the same sentence 'A dog scared every cat' appears with the dog as x/3 in Section 3.4 and as x/1 in Section 3.5. In the second version the inferred context 3 is empty, so the DRS that results puts the existential dog inside the consequent of the universal, giving 'every cat was scared by some dog' rather than the wide-scope reading intended in Section 3.4. That is not a mere typo: it shows the mapping from indices to contexts is not deterministic, and the claimed equivalence with PMB DRS is not yet shown.\n\nThe good news is the flaws are fixable. A revision could give a recursive semantics for 'l : C', define τ properly, and redo the example consistently. The paper already acknowledges related work, including named graphs, and the idea of monotonically extending existing AMR annotations is practical.\n\nI would not cite this as a formal result. I might cite it as a pointer to the general approach. It deserves a serious referee, because the problem is real and the proposal is a reasonable direction, but the manuscript as it stands needs major revision before the central claim can be accepted. If it lands on my desk, I would invite a revision rather than accept or reject outright.","headline":"The AMR+ proposal is a clean idea, but the paper's central equivalence claim rests on a translation that is both underspecified and internally inconsistent in the headline example.","tokens_in":6968,"tokens_out":5956,"would_cite":false,"duration_ms":57102,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AMR can be extended with a logical layer of indexed contexts that handles negation, quantifier scope, and bound variables.","keywords":["Abstract Meaning Representation","AMR+","scope","quantification","negation","presupposition","Discourse Representation Theory","context indexing"],"falsifier":"Apply the paper's three-step translation to \"A dog scared every cat\" exactly as in Section 3.5: the resulting DRS is ∀y(cat(y)→∃x(dog(x)∧scare(x,y))), which is not the surface-scope reading ∃x(dog(x)∧∀y(cat(y)→scare(x,y))). If the intended AMR+ for that sentence admits only the weaker reading, then the formalism under-generates; if both readings are intended, the constraints lack a mechanism for deriving the missing one. Either way, checking which reading the notation actually delivers would settle the claim that AMR+ correctly handles quantifier scope.","tokens_in":5946,"feed_emoji":"🧩","tokens_out":7092,"duration_ms":73650,"temperature":0.7,"pith_summary":"The paper argues that Abstract Meaning Representation, which excels at predicate-argument structure, can be given a logical layer without losing that structure. By labeling each sub-AMR with an index and adding constraints on how those contexts relate, the paper derives a scoped formalism it calls AMR+. This is claimed to solve two known problems: the bound variable problem, where AMR graphs misread \"every snake bit itself,\" and the scope representation problem, where AMR's predicate-style negation permits invalid inferences such as concluding \"Mary left\" from \"Mary did not leave.\" The paper further claims AMR+ is equivalent in expressive power to the DRS dialect used in the Parallel Meaning Bank, which would transfer DRT's inferential properties to AMR.","feed_headline":"A logical layer lets AMR handle quantifier scope","feed_subtitle":"By indexing sub-AMRs, negation, quantification, and presupposition gain explicit scope without changing argument structure.","key_machinery":"The central object is the indexed context layer: every slash in a PENMAN AMR is decorated with a numeric index, and a set of constraints over those indices describes the logical relationships between the corresponding sub-AMRs. The constraints use identity, negation, conditional implication, and presupposition (‘=’, ‘¬’, ‘⇒’, ‘<’), with colons expressing that a context contains a condition. This machinery recursively decomposes an AMR into sub-AMRs and then reassembles them into a scoped structure, separating logical structure from the surface-friendly argument structure that AMR is known for.","core_discovery":"The central claim is that adding a second, logical layer to AMRs—indices on every sub-AMR plus constraints stating how contexts relate as identical, negated, conditional, or presuppositional—makes AMR+ correct for inferences involving negation and bound variables while preserving the original predicate-argument structure. Negation becomes a scoping operator rather than a predicate attribute, universal quantification becomes a conditional context, and definite descriptions and names trigger presuppositional contexts. A three-step translation converts AMR+ into DRS: turn each sub-AMR into a DRS, merge DRSs with the same index, and combine them by following the constraints. On the paper's account, AMR+ is equivalent in expressive power to the Parallel Meaning Bank's DRS dialect, and this is what would give AMR the inferential behavior of Discourse Representation Theory.","pith_inferences":["Taken literally, the paper's own translation of \"A dog scared every cat\" in Section 3.5 yields the reading ∀y(cat(y)→∃x(dog(x)∧scare(x,y))), with \"every cat\" wide over \"a dog,\" rather than the surface reading ∃x(dog(x)∧∀y(cat(y)→scare(x,y))); the paper does not flag this discrepancy, so the intended scope readings are not yet fully pinned down.","A model-theoretic semantics for the context constraints themselves—rather than only for their DRS translations—would make the claimed inference predictions checkable and would settle whether the constraints are compositional.","The constraint vocabulary looks portable: other graph-based meaning representations that suffer from the same bound-variable problem could adopt indexed-context constraints, though the paper does not claim this.","The proposal implicitly treats quantifier scope as a property of the representation rather than of the parsing process; testing it on genuinely ambiguous sentences would clarify whether AMR+ stores one reading per structure or can encode multiple readings."],"forward_implications":["Existing AMR-annotated corpora can be extended monotonically: add indices automatically by replacing every \"/\" with \"/1/\", then manually fix cases of negation, universal quantification, and definite descriptions.","The \":polarity\" and \":mod\" relations in AMR become optional or removable, because the new context constraints take over their logical work.","Because AMR+ can be translated into the Parallel Meaning Bank's DRS dialect through the three-step recipe, DRT's model-theoretic components and inference procedures could be applied to AMR.","The paper's formalism gives AMR+ the same expressive power as the Parallel Meaning Bank's DRS dialect, though not the full expressive power of general DRS.","A modular parsing strategy becomes possible: existing AMR parsers can remain unchanged, with a separate layer adding the logical constraints."],"supporting_citations":[{"why":"Supplies the AMR-annotated corpus that the extension is designed to build on and monotonically extend.","marker":"(Banarescu et al., 2013)"},{"why":"Provides the PENMAN slash notation whose sub-AMRs receive the new indices.","marker":"(Kasper, 1989)"},{"why":"Introduces the AMR formalism as the representational starting point of the paper.","marker":"(Langkilde and Knight, 1998)"},{"why":"Earlier reformulation of AMR as a recursive structure that the indexing operation relies on.","marker":"(Bos, 2016)"},{"why":"Prior attempt to reform AMR for scope that this proposal claims to improve on in simplicity.","marker":"(Stabler, 2017)"},{"why":"Defines DRT and DRS, the framework whose inferential properties AMR+ aims to inherit.","marker":"(Kamp and Reyle, 1993)"},{"why":"Defines the Parallel Meaning Bank's DRS dialect, to which AMR+ is claimed to be equivalent in expressive power.","marker":"(Abzianidze et al., 2017)"},{"why":"Provides the presupposition-as-anaphora-resolution treatment that inspires the presuppositional context constraint.","marker":"(Van der Sandt, 1992)"}],"fun_headline_variants":["AMR gains quantifier scope with added logical layer","Indexing sub-AMRs gives AMR DRT-style inference","Separating argument and logical structure in AMR","AMR's logical layer adds explicit quantifier scope","AMR + indices gives DRT-like inferential power"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim stands on the assumption that the informally specified translation from AMR+ context constraints to DRS faithfully captures the meaning of those constraints; no model-theoretic semantics is given, and the paper's own example \"A dog scared every cat\" is indexed differently in Section 3.4 than in Section 3.5.","fun_headline_variants_meta":{"raw":{"variants":["AMR gains quantifier scope with added logical layer","Indexing sub-AMRs gives AMR DRT-style inference","Separating argument and logical structure in AMR","AMR's logical layer adds explicit quantifier scope","AMR + indices gives DRT-like inferential power"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":2873,"prompt_tokens":764,"completion_tokens":2109,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":380,"completion_tokens_details":{"reasoning_tokens":2030}},"tokens_in":380,"tokens_out":2109,"duration_ms":14917,"temperature":1.0,"reasoning_tokens":2030,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:15:16.261967+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the paper's three-step translation to \"A dog scared every cat\" exactly as in Section 3.5: the resulting DRS is ∀y(cat(y)→∃x(dog(x)∧scare(x,y))), which is not the surface-scope reading ∃x(dog(x)∧∀y(cat(y)→scare(x,y))). If the intended AMR+ for that sentence admits only the weaker reading, then the formalism under-generates; if both readings are intended, the constraints lack a mechanism for deriving the missing one. Either way, checking which reading the notation actually delivers would settle the claim that AMR+ correctly handles quantifier scope.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the AMR-annotated corpus that the extension is designed to build on and monotonically extend."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the AMR formalism as the representational starting point of the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier reformulation of AMR as a recursive structure that the indexing operation relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior attempt to reform AMR for scope that this proposal claims to improve on in simplicity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines DRT and DRS, the framework whose inferential properties AMR+ aims to inherit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Parallel Meaning Bank's DRS dialect, to which AMR+ is claimed to be equivalent in expressive power."}],"review_version":1}