{"id":"5ec3458d-1ad5-4338-845e-d827a8d3db6d","arxiv_id":"2607.06542","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Fast-decaying sequence-length distributions make random free-tree parsers recover high edge accuracy on primate sequences, rendering gold-standard-free evaluation feasible unlike in human language.","lead":"Random dependency parsers recover a high fraction of correct edges on non-human primate vocal and gestural sequences because those sequences are short and their length distributions decay exponentially. This supplies a usable lower bound on accuracy without any gold-standard treebank, something human languages lack.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The central mathematical derivation (expected edge accuracy 2/n for a random free tree, then expectation under geometric or empirical length distributions) is standard combinatorics and checks out against the published tables. The only substantive modeling premise—the unique equiprobable free tree—is already flagged by the reader and is the reason for CONDITIONAL rather than ACCEPT. No further soft spot (data leakage, circularity, or numerical error) appears that would move the verdict. The suggested concrete test simply quantifies sensitivity to that already-identified premise; a large drop would confirm the reader's caution, while stability would strengthen the claim without changing the present verdict.","tokens_in":31068,"tokens_out":403,"duration_ms":5326,"concrete_test":"Recompute E[Q] and E[P^e_c] for the empirical chimpanzee and gelada length histograms after replacing the uniform-tree prior with a degree-biased prior (e.g., preferential-attachment spanning trees); if the resulting lower bounds remain >50 % for n_min=2, the feasibility claim is robust to the equiprobability assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest-assumption note (unique equiprobable free tree) is already the paper's own explicit modeling choice (§2.1, Appendices B–C). It is used only to obtain a lower bound on random-parser edge accuracy via Cayley's formula and the London–Pluhár intersection expectation; the subsequent claim that any good-enough parser must exceed that bound therefore remains valid under the stated premises. No internal inconsistency, hidden parameter, or data-selection artifact undermines the numerical results for geladas, chimpanzees or the 31-species ensemble. The modeling choice is untested for non-human sequences, but that is already reflected in the CONDITIONAL verdict and does not constitute an additional load-bearing flaw.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper claims that unsupervised free-tree dependency parsing of non-human primate vocal/gestural sequences can be evaluated without a gold standard. Because sequence lengths are short and decay rapidly (geometric-like), even a random parser that samples uniformly from Cayley’s labeled trees already recovers a high expected fraction of correct edges (E[Q] or E[P^e_c]). Any statistically informed (“good-enough”) parser therefore inherits a high lower bound on accuracy. Analytic expressions for E[P^t_c], E[Q] and E[P^e_c] are derived under uniform, geometric and empirical length distributions (Appendices B–E, using the London–Pluhár intersection formula). Empirical estimates reach 51 % (geladas) and >79 % (chimpanzees) for n_min=2, versus ~13–28 % on human PUD sentences; the same pattern holds for a 31-species ensemble when only n_max is known. Hence evaluation is feasible for primates but hard for human languages.","tokens_in":31208,"tokens_out":916,"duration_ms":31429,"significance":"If the result holds, it removes a long-standing methodological barrier to applying dependency-parsing tools outside human language and supplies an immediately usable, parameter-free lower-bound protocol. Strengths include fully analytic derivations controlled to 10^{-8} error, transparent use of published multi-species length data, parallel human baselines (PUD/PUD10), and an explicit end-to-end methodology (§5.3). The work therefore has clear value for both computational linguistics and comparative communication research.","major_comments":[{"comment":"The lower-bound argument (abstract, §2.3–2.6, §5.1) is sound under the stated premises, yet the abstract and conclusion phrase the result as an unconditional “must be high.” The claim is conditional on every sequence possessing a unique correct free tree (explicitly assumed in §2.1 and used to invoke Cayley + London–Pluhár in Appendices B–C). A brief paragraph acknowledging that the numerical guarantees collapse if animal sequences lack a unique tree (or are better modeled as forests/DAGs) would prevent over-reading.","section":"abstract, §2.1, Appendices B–C, §5.1"},{"comment":"Table 4 and the 31-species analysis rely on the uniform distribution as a lower bound for any non-increasing length distribution. While this is mathematically correct, the paper never verifies that the length distributions of the 29 species for which only n_max is known are in fact non-increasing. A single sentence noting that the bound is conservative only under that additional empirical premise would tighten the claim.","section":"§4.2, Table 4"}],"minor_comments":[{"comment":"Introduction contains multiple missing spaces (“asyntactic”, “atreebank”, “agold standard”, “Unsupervised parserslearn”).","section":"§1"},{"comment":"Placeholder text remains: “Action editor: {action editor name}”, “Submission received: DD Month YYYY”.","section":"front matter"},{"comment":"Figure 4 caption refers to a “2-parameter geometric distribution” while the plotted curve uses the MVU estimator of Appendix A; a one-line clarification would help.","section":"Figure 4"},{"comment":"Tables 5–8 and G.1–G.2 are dense; adding a short note that all geometric expectations use ε=10^{-8} (Appendix E) would improve reproducibility.","section":"Tables 5–8, Appendix G"},{"comment":"Eq. (19) defines E[Q]^* but is never used numerically; either drop it or show the comparison for the empirical cases.","section":"§2.3"}],"recommendation":"minor_revision","confidential_remarks":"Solid short paper that fits Computational Linguistics. The modeling assumptions are transparent and the mathematics clean; the two major comments are easily addressable. No novelty or citation concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: because primate vocal and gestural sequences are short and their length distributions decay fast (geometric-like), a random free-tree parser already recovers a large fraction of edges (roughly 50–80 % depending on n_min and species). Any parser that beats chance therefore has a high lower bound on accuracy even without a gold standard. Humans lack that property; the same baseline sits around 13–28 %. That is the concrete methodological claim.\n\nWhat is new is the application. The math itself is classical (Cayley’s formula for labeled trees, London & Pluhár’s expected intersection of two random spanning trees). The authors derive E[P^t_c], E[Q] and E[P^e_c] cleanly in the appendices, control the geometric truncation to 10^{-8}, and then plug in published length counts for geladas, chimpanzees (four gesture criteria) and a 31-species n_max table, plus parallel PUD for humans. The tables are transparent and the contrast is sharp. They also sketch a practical protocol (check statistical structure, check fast decay, train, validate against shuffled baseline, report the random lower bound). That is useful for anyone working on animal sequences.\n\nThe soft spot is exactly the modeling premise they state up front: every sequence has one correct free (undirected) tree and all labeled free trees of size n are equiprobable. That is standard for human dependency evaluation, but untested for other species. If the true structure is non-unique or heavily non-uniform, the numerical lower bounds no longer apply. The paper is open about this; it does not hide it. Everything else (data provenance, numerical stability, citation of the network-science source) looks solid. Self-citations are to earlier empirical work on the same sequences and are not load-bearing for the combinatorics.\n\nThis is for people who actually want to run unsupervised parsers on animal data or who care about quantitative ethology. It will not reorganize core linguistics, but it removes a concrete roadblock. I would send it to referees; the math and data are clear enough that a specialist can decide whether the free-tree axiom is acceptable for the target species. Worth engaging if that is your area.","headline":"Clean combinatorial lower bounds show random free-tree parsers already hit high edge accuracy on short primate sequences, making evaluation without gold standard feasible there but hard for humans.","tokens_in":31818,"tokens_out":549,"would_cite":true,"duration_ms":7097,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Short primate sequences make random dependency parsing accurate enough that evaluation without gold trees becomes feasible.","keywords":["dependency parsing","unsupervised parsing","evaluation without gold standard","sequence length distribution","non-human primate communication","free trees","network science"],"falsifier":"Collect a new primate corpus whose length distribution is no longer short-tailed (or whose true trees are known by independent means) and check whether a random free-tree parser still recovers more than half the edges; if accuracy collapses or the true trees systematically deviate from the uniform model, the lower-bound argument fails.","tokens_in":31947,"feed_emoji":"🐒","tokens_out":816,"duration_ms":9157,"temperature":0.7,"pith_summary":"The paper asks whether one can evaluate an unsupervised dependency parser on animal vocal or gestural sequences when no gold-standard trees exist. It shows that a random free-tree parser already recovers a large fraction of correct undirected edges simply because those sequences are short and their length distribution decays exponentially. For gelada and chimpanzee data the expected accuracy sits between 50 % and 80 %; for human sentences of unrestricted length it falls to roughly 13 %. Consequently any parser that beats chance is guaranteed a high lower bound on accuracy in the animal case, while the same guarantee is unavailable for ordinary human text. The result turns the usual pessimism about non-human syntax on its head: evaluation without gold data is actually easier for other primates than for us.","feed_headline":"Short primate sequences make random parsing surprisingly accurate","feed_subtitle":"No gold trees needed: length decay alone guarantees high edge accuracy for geladas and chimps","key_machinery":"The London–Pluhár expectation that two uniform random labeled free trees on n vertices share 2(n−1)/n edges; averaging this quantity over the empirical or geometric length distribution of a corpus yields the lower bounds E[Q] and E[P_e^c].","core_discovery":"Because non-human primate sequence lengths decay rapidly (often geometrically), the expected fraction of correct undirected edges recovered by a random free-tree parser is high—above 50 % for geladas and above 79 % for chimpanzees—while the same quantity for unrestricted human sentences is only about 13 %. Therefore a lower bound on parser accuracy can be stated without any gold standard for those species, rendering evaluation feasible where it is hard for human language.","pith_inferences":["The argument supplies a practical checklist—fast length decay plus some statistical structure—for deciding whether unsupervised parsing is worth attempting on a new animal communication system.","If the uniform-tree prior is replaced by a more concentrated distribution over trees, the lower bounds would only improve, so the feasibility claim is conservative.","The same length-driven accuracy effect may apply to other short-sequence domains (e.g., short social-media posts or early child language) that have been treated as hard for unsupervised parsing."],"forward_implications":["Any statistically informed unsupervised parser trained on gelada or chimpanzee sequences is guaranteed a high lower bound on undirected edge accuracy even without gold data.","The same guarantee does not hold for ordinary human sentences, so unsupervised evaluation remains hard for human languages.","Curriculum-style training that begins with short sequences is automatically favored by the natural length distribution of primate data.","The same network-science calculation can be applied to any other species once its sequence-length histogram is known."],"fun_headline_variants":["Primate length decay guarantees high random-parse edge accuracy","Short sequences let us evaluate unsupervised parsers without gold trees","Random free-tree accuracy exceeds 50% for geladas, 79% for chimps","Fast decay of vocalization lengths enables gold-free parser evaluation","Unrestricted human sentences lack the accuracy floor of primate data"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Every sequence is assumed to possess exactly one correct free tree and all labeled free trees of the same size are equally likely a priori.","fun_headline_variants_meta":{"raw":{"variants":["Primate length decay guarantees high random-parse edge accuracy","Short sequences let us evaluate unsupervised parsers without gold trees","Random free-tree accuracy exceeds 50% for geladas, 79% for chimps","Fast decay of vocalization lengths enables gold-free parser evaluation","Unrestricted human sentences lack the accuracy floor of primate data"]},"model":"grok-4.5","effort":"low","cost_usd":0.003058,"raw_usage":{"total_tokens":1019,"prompt_tokens":716,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":30580000,"prompt_tokens_details":{"text_tokens":716,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":231,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":716,"tokens_out":72,"duration_ms":4090,"temperature":1.0,"reasoning_tokens":231,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T00:05:29.282350+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Collect a new primate corpus whose length distribution is no longer short-tailed (or whose true trees are known by independent means) and check whether a random free-tree parser still recovers more than half the edges; if accuracy collapses or the true trees systematically deviate from the uniform model, the lower-bound argument fails.","supporting_citations":[],"review_version":2}