{"id":"0dd79caa-f0d2-42f7-90fd-fd0d80122399","arxiv_id":"2607.10591","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"Arity is not a necessary transition primitive for non-binary bottom-up constituency parsing; delimiter-bounded stack configurations uniquely determine reduction spans.","lead":"Non-binary bottom-up parsers can drop REDUCE-X#k actions that name both a label and its child count. Stack delimiters make the child count recoverable from parser state, yielding a smaller action set that stays competitive on PTB and CTB.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The §3.3 static oracle for D-SHIFT fails on post-nested siblings, so the written rules cannot reconstruct many gold trees despite strong reported F1.","rationale":"The reader’s weakest assumption (learnability of an already-correct oracle) presupposes that the §3.3 insertion rule produces gold-compatible sequences. The rule as written does not; the mismatch is a prior correctness gap in the method that must be resolved (most likely the implementation used the proper “left-edge of any phrasal node” test while the text gave an incomplete local approximation). Once the oracle is stated correctly the formal recoverability claim and the empirical no-collapse results stand, so the paper remains publishable after a clarification, hence CONDITIONAL rather than REJECT. The reader’s ACCEPT is therefore too strong given the specification error, but the underlying idea is intact.","tokens_in":14430,"tokens_out":617,"duration_ms":116594,"concrete_test":"Build the gold tree for a minimal nested-subject sentence (e.g., “The man who runs plays .”). Apply the exact D-SHIFT rule of §3.3 to every terminal, then execute the full static-oracle sequence of NT/U-REDUCE/REDUCE/D-REDUCE. If the derived tree differs from gold (VP or S will have wrong children/arity), the written oracle is unsound and the method description does not support the experiments.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim relies on a static oracle that inserts delimiters so every gold phrase-level constituent has a recoverable delimiter-bounded domain (then REDUCE/D-REDUCE according to the highest associated ancestor a(D)). Section 3.3 defines D-SHIFT(wi) = 1 only on sentence start or when g(wi) ≠ g(wi-1) and d(wi)-d(wi-1) ≥ 0 (g = grandparent/phrase projection, d = terminal depth). This rule correctly handles flat same-depth transitions and left-aligned nesting, but suppresses the required D-SHIFT whenever a new sibling phrase begins after a nested previous sibling: the last terminal of the nested material is deeper, so the depth difference is negative even though g changes and the new phrase’s left edge must be marked. Without that delimiter the subsequent REDUCE for the sibling (e.g., VP after a complex NP) sees only the outer delimiter and incorrectly consumes preceding siblings, producing a non-gold tree. Because PTB/CTB contain abundant nested subjects, PPs, etc., the published oracle cannot generate correct gold derivations for a large fraction of trees; yet Tables 1–5 report high F1 and near-gold arity profiles. The recoverability argument (well-formed configurations) remains definitionally true, but the concrete transition system that is claimed to realize it is incompletely/incorrectly specified.","agreement_with_reader":"disagree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a non-binary bottom-up transition system for constituency parsing in which reduction arity is not an action parameter. Constituent labels are introduced by NT(X), while reduction spans are recovered from dummy delimiters placed by D-SHIFT and managed by REDUCE vs. D-REDUCE. In a well-formed configuration σ|D|s1|…|sk|X, arity k is uniquely determined by the stack interval between the active delimiter and the label marker. The system is evaluated against a reimplemented arity-specific baseline (REDUCE-X#k) and a binary bottom-up baseline in a shared Stanza framework on PTB and CTB, reporting competitive F1, a substantially smaller action inventory, near-gold arity distributions, and no high-arity collapse.","tokens_in":14833,"tokens_out":1091,"duration_ms":39913,"significance":"If the transition system is correctly specified and the empirical results hold under that specification, the paper makes a clean and useful design contribution: it shows that direct non-binary bottom-up construction need not encode the cross-product of labels and arities in the action inventory. The controlled same-framework comparison, inventory compression (Table 3), arity-profile analysis (Tables 4–5), and explicit boundary-error audit (Table 7) are strengths. The contribution is incremental rather than transformative, but it reframes bottom-up reduction as a state-recoverable domain problem and is of clear interest to transition-based parsing.","major_comments":[{"comment":"Section 3.3 (D-SHIFT oracle): the stated rule D-SHIFT(wi)=1 only when i=0 or when g(wi)≠g(wi-1) and d(wi)-d(wi-1)≥0 fails for a common gold configuration—post-nested siblings. After a nested child (e.g., a deep PP inside a subject NP), the next sibling phrase (e.g., VP) has a changed grandparent but strictly smaller terminal depth, so the rule suppresses the required left-edge delimiter. Without that delimiter, the subsequent phrase-level REDUCE over the outer active D incorrectly consumes the preceding sibling(s). PTB/CTB contain many such cases; the written static oracle therefore cannot generate correct gold derivations for a large fraction of trees, even though recoverability is definitional for well-formed configurations (Section 3.5). This is load-bearing for the claim that the published system constructs original non-binary trees. Please give a corrected, complete oracle (with a n","section":null},{"comment":"Appendix A / relation of oracle to association rule a(D): Section 3.3 associates each delimiter with the highest phrasal ancestor whose left edge coincides with the shifted token, and chooses REDUCE vs. D-REDUCE by whether the completed constituent equals a(D). The incomplete D-SHIFT rule above breaks this association for right siblings after nesting. Even if the depth condition is only a sketch, the paper currently provides no alternative decision procedure that is both (i) local enough for a static oracle and (ii) complete for gold trees. A short completeness argument or oracle pseudocode covering nested and flat branching is needed before the factorization claim can be assessed as realized by the concrete system.","section":null}],"minor_comments":[{"comment":"Table 2: the upper block mixes published numbers from heterogeneous setups with the controlled lower block; a clearer visual separation or caption note would avoid over-reading cross-paper F1 gaps.","section":null},{"comment":"Table 1 vs. Table 2: development binary F1 exceeds both non-binary systems by a large margin on CTB; a brief discussion of why binarization helps under this encoder would help readers interpret the non-binary comparison.","section":null},{"comment":"Section 5.5 / Table 7: error counts are absolute; rates per sentence or per gold delimiter would make PTB vs. CTB severity easier to compare.","section":null},{"comment":"Figure 1 and Section 3.2: FINISH is listed but the completion flag f is otherwise unused in the transition semantics; a one-line note on why f is retained (framework compatibility) would avoid confusion.","section":null},{"comment":"Notation: D is introduced as a dummy delimiter implemented as empty NT(); stating once whether NT() is a distinct action or only an internal stack token would clarify the inventory in Table 3.","section":null}],"recommendation":"major_revision","confidential_remarks":"The empirical tables strongly suggest a working implementation, so the most likely issue is an incomplete oracle write-up rather than fabricated results. I would not reject on that basis, but I would not accept until the oracle is corrected and shown complete for nested siblings. If the authors cannot produce a correct static oracle matching the reported runs, the central claim would need to be re-scoped."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: you can drop REDUCE-X#k entirely from non-binary bottom-up constituency parsing. Introduce the label with NT(X), mark left edges with a dummy delimiter via D-SHIFT, and let REDUCE/D-REDUCE recover the child span from the stack interval. Arity becomes a derived property of a well-formed configuration rather than a classifier target. That factorization is new relative to Fernández-González & Gómez-Rodríguez 2019 and keeps direct construction of the original trees.\n\nThey run the comparison cleanly. Same Stanza encoder, composition, optimizer, and eval script for the binary baseline, the arity-specific reimplementation, and their delimiter system. Inventory shrinks from ~200 to ~80 actions, development and test F1 stay competitive with the arity baseline, the predicted arity histogram tracks gold, and high-arity F-scores do not collapse. They also document the real costs—longer transition sequences and the new error modes around delimiter insertion and persistence—without overselling. That is useful, proportionate work.\n\nThe soft spot is the static oracle in §3.3. The printed rule (D-SHIFT on sentence start or grandparent change with non-negative depth delta) correctly handles flat same-depth siblings and entry into nests, but it suppresses the required delimiter when a new sibling phrase begins after nested material: depth decreases, so the condition fails even though the grandparent changes. Complex subjects followed by VP are common in PTB/CTB; without that delimiter the subsequent reduce would pull in the wrong material. Yet the reported F1 and arity tables look solid. The recoverability claim itself is definitional and fine; the concrete oracle description is incomplete or imprecise. Probably the implementation marks left edges correctly and the text is just underspecified, but it needs fixing.\n\nEverything else is minor: numbers are not SOTA, only two treebanks, longer derivations, no public code mentioned. Citations are standard and fair. This is for people who design transition systems and care about action inventories. The core methodological result is real and well-supported under the matched setup. Send it to referees; it deserves the discussion.","headline":"Clean factorization showing arity need not be an action label in non-binary bottom-up parsing; controlled reimpl holds up, but the printed D-SHIFT oracle rule is incomplete for common nested-then-sibling cases.","tokens_in":15374,"tokens_out":551,"would_cite":true,"duration_ms":45358,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Arity is not a necessary transition primitive for non-binary bottom-up constituency parsing.","keywords":["constituency parsing","transition-based parsing","non-binary trees","bottom-up parsing","arity","stack delimiters","action inventory"],"falsifier":"On the same PTB and CTB splits and framework, show that the delimiter-guided parser’s high-arity (k≥4 or k≥5) constituent F-scores fall sharply relative to the arity-specific baseline, or that its predicted arity histogram diverges strongly from gold while overall F1 remains high only by collapsing to low-arity trees.","tokens_in":15324,"feed_emoji":"🌳","tokens_out":588,"duration_ms":9804,"temperature":0.7,"pith_summary":"Bottom-up constituency parsers that build original non-binary trees usually pack the mother label and the number of children into one action, such as Reduce-X#k. This paper argues that the arity number k is not required as an action at all. Constituent labels are introduced separately, while stack delimiters mark the left edge of each reduction domain so that the child span is read off the parser state. In a well-formed configuration the arity is uniquely fixed by the active delimiter and the label marker, so it becomes a derived property rather than a classifier target. On standard English and Chinese treebanks the resulting system stays competitive with an arity-specific baseline under the same implementation, uses a much smaller action inventory, and does not collapse high-arity constituents.","feed_headline":"Arity need not be an action in bottom-up parsing","feed_subtitle":"Stack delimiters recover child spans so Reduce-X#k actions can be dropped without collapsing high-arity trees","key_machinery":"Delimiter-guided reduction (arity recoverability): dummy delimiters placed by D-SHIFT bound the stack interval that REDUCE or D-REDUCE will compose under a separately introduced label, so k is an invariant of state rather than an action label.","core_discovery":"The arity parameter of non-binary bottom-up reduction is not a necessary transition primitive. Once labels are introduced by a separate NT(X) action and reduction domains are bounded by stack delimiters, arity is uniquely recovered from the completed items between the active delimiter and the label marker, while still constructing original non-binary trees directly.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Arity recovered from stack delimiters, not reduce actions","Non-binary bottom-up parsing drops arity from reduce labels","Delimiter-bounded stacks make arity a derived property","Label first, recover spans: no arity-specific reduce needed","High-arity trees stay intact without arity transition actions"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The training oracle’s rule for when to insert and when to keep or consume delimiters must produce sequences a neural policy can learn well enough that recovered spans stay accurate; if boundary decisions systematically fail, arities go wrong even though recoverability holds by construction.","fun_headline_variants_meta":{"raw":{"variants":["Arity recovered from stack delimiters, not reduce actions","Non-binary bottom-up parsing drops arity from reduce labels","Delimiter-bounded stacks make arity a derived property","Label first, recover spans: no arity-specific reduce needed","High-arity trees stay intact without arity transition actions"]},"model":"grok-4.5","effort":"low","cost_usd":0.002906,"raw_usage":{"total_tokens":1021,"prompt_tokens":709,"num_sources_used":0,"completion_tokens":81,"cost_in_usd_ticks":29060000,"prompt_tokens_details":{"text_tokens":709,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":231,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":709,"tokens_out":81,"duration_ms":3036,"temperature":1.0,"reasoning_tokens":231,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T10:36:01.905636+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same PTB and CTB splits and framework, show that the delimiter-guided parser’s high-arity (k≥4 or k≥5) constituent F-scores fall sharply relative to the arity-specific baseline, or that its predicted arity histogram diverges strongly from gold while overall F1 remains high only by collapsing to low-arity trees.","supporting_citations":[],"review_version":1}