{"id":"afa3dd99-b350-44d2-aae6-8b26095db15b","arxiv_id":"2607.03681","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"ETM is a shared morphological exponent of multiple adnominal constructions; relative-clause-like uses are only 39.4% of KLUE training instances and cannot be identified from the ending alone.","lead":"Korean ETM morphology marks many noun-modifying constructions, not only relative clauses; only 39.4% of KLUE instances are productive relative-clause-like uses. The paper supplies a rule-based, construction-sensitive annotation layer that separates adjectival, copular, bound-noun, modal, temporal and collocational patterns.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The reader correctly isolates the strongest empirical claim and the most plausible soft spot (rule-order fidelity on borderline cases). That soft spot is real but not load-bearing: even if every jointly rejected and uncertain validation case were reassigned from non-RC to RC, the RC share would rise only a few points and would still leave non-RC constructions as the majority. The paper already surfaces the same boundary problems in §5 and Table 2, retains an explicit undefined residual, and ships the code and schema needed for external re-labeling. Because the quantitative claim survives reasonable reclassification noise and no deeper methodological flaw is present, the CONDITIONAL verdict stands without adjustment.","tokens_in":20710,"tokens_out":434,"duration_ms":3955,"concrete_test":"Independently re-run Algorithm 1 on a fresh random sample of 200 ETM tokens drawn from the same KLUE training split (stratified by the 11 defined labels), re-label them manually under the published schema, and recompute the RC-like share; if the new share remains within 39.4%±5 pp the headline claim is stable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (ETM is shared morphology; productive RC-like uses are only 39.4% of 13 046 KLUE training instances) is supported by a fully specified ordered rule procedure (Algorithm 1), explicit category inventory (Table 1), released reconstruction artifacts, and a stratified 85-instance manual check (91.8% agreement, κ=0.677, 90.9% joint acceptance among definite judgments). The residual boundary issues noted by the reader (bound-noun vs. RC, predicate-class ambiguity) are already quantified as small (undefined=0.6%; jointly rejected cases=7/85) and do not reverse the majority non-RC distribution. No internal inconsistency or circularity appears; the result is a transparent corpus-engineering finding rather than a fragile theoretical leap.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper argues that the Korean adnominal ending ETM is not a direct marker of relative-clause structure but a morphological form shared by several noun-modifying constructions. It proposes a constructional typology based on predicate type, auxiliary structure, head-noun restriction, argument-structural compatibility, and lexicalized patterns, operationalized as an ordered rule-based annotation layer over the KLUE dependency treebank (13,046 ETM instances in the training split). Manual validation on 85 stratified instances yields 91.8% observed agreement (κ = 0.677) and 90.9% joint acceptance among definite judgments. Corpus analysis finds that productive relative-clause-like uses account for only 39.4% of instances; the remainder is mainly adjectival, copular, bound-nominal, modal, temporal, and collocational. The authors release the label schema, code, documentation, and derived labels.","tokens_in":20887,"tokens_out":1164,"duration_ms":17910,"significance":"If the result holds, the paper supplies a concrete, reproducible correction to a common practice in Korean corpus work and treebanking: treating ETM (or coarse labels such as vp_mod) as a proxy for relative clauses. The finding that non-RC constructions form the majority of ETM tokens is directly useful for annotation guidelines, UD mapping debates, and any downstream study that searches for Korean relative clauses by morphology. Strengths include a fully specified decision procedure (Algorithm 1), an explicit category inventory (Table 1), transparent residual handling (undefined = 0.6%), released reconstruction artifacts under the source license, and quantified manual validation with confidence intervals. This is solid corpus-engineering work with clear linguistic motivation rather than a fragile theoretical leap.","major_comments":[{"comment":"§5 (manual validation) and Table 3: The central quantitative claim (39.4% RC-like) rests on the ordered rules being correct at scale, but validation covers only 85 stratified instances and checks accept/reject of automatic labels rather than independent double annotation of construction types. κ = 0.677 is moderate, and the authors themselves locate residual error at the bound-noun vs. RC boundary (modif+etm+nnb = 2,774 vs. rc+vv+etm = 4,816). A category-wise confusion summary or a larger gold sample focused on that boundary would make the reported percentage more robust; without it, the precise 39.4% figure carries residual uncertainty even though the majority non-RC conclusion is unlikely to reverse.","section":"§5 Manual validation; Table 3"},{"comment":"§4.3 and Algorithm 1: The linguistic precedence among construction types is well motivated, but the paper reports no sensitivity analysis on order or on the fixed collocation list (Appendix A.1 / Lim 2009). Because several large classes (especially modif+etm+nnb and colloc+etm) are resolved before productive RC categories, modest reordering or list incompleteness could shift counts between non-RC and RC buckets. A short ablation (e.g., bound-noun after RC; collocation list variants) would show how stable the 39.4% figure is under reasonable alternative precedence choices.","section":"§4.3 Decision procedure; Algorithm 1"}],"minor_comments":[{"comment":"§3.1 / Figure 1: The example tree is helpful, but the romanization and gloss lines are dense; a slightly larger font or a simplified dependency excerpt focused on the ETM token (높+은) would improve readability in print.","section":"§3.1 Figure 1"},{"comment":"Table 1 and §4.2: The label names (e.g., modif+su-iss+etm, rc+va+vx+etm) are transparent once defined but heavy in running text; a short mnemonic key or consistent English glosses in the table would help non-Koreanist readers.","section":"Table 1; §4.2"},{"comment":"§6: The overgeneration calculation (7,911 / 13,046 = 60.6%) is clear; stating explicitly that this is under a naïve “all ETM = RC” baseline would prevent misreading it as a parser error rate.","section":"§6 Relative clause and non-relative uses"},{"comment":"References: Several Korean-language sources are appropriately cited; ensuring consistent romanization of author names and journal titles across the list would aid indexing.","section":"References"},{"comment":"Appendix A: The examples are excellent; numbering them continuously with the main text or cross-referencing category counts from Table 3 in each subsection would tighten the link between typology and distribution.","section":"Appendix A"}],"recommendation":"minor_revision","confidential_remarks":"Fit for a computational linguistics / corpus linguistics venue is good: the contribution is annotation methodology plus a clear empirical distribution, not a new parsing model. Novelty is incremental but well executed; the main risk is that reviewers may undervalue rule-based annotation layers relative to neural tagging papers. The validation sample is the only point I would press the authors on before acceptance; I would not require a second corpus unless the journal expects multi-resource evaluation."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is the clean empirical result: of 13,046 ETM tokens in the KLUE training split, only 39.4% are productive relative-clause-like; the rest are adjectival, copular, bound-noun, modal, temporal, and collocational constructions. Morphology alone does not identify Korean relative clauses.\n\nWhat is new is the operationalization. The authors pull together diagnostics that Korean linguists already knew (predicate type, head-noun restriction, auxiliaries, lexicalization) into one ordered rule procedure (Algorithm 1), apply it exhaustively, and ship the label schema, code, and full derived layer so anyone can reconstruct it. Table 3 and the head-noun patterns in Table 4 make the distribution concrete. Manual check on 85 stratified items gives 91.8% agreement and κ=0.677; among definite judgments the automatic label is accepted 90.9% of the time. Residual undefined is only 0.6%. That is honest engineering, not circular counting.\n\nSoft spots are real but limited. The validation sample is small, and the authors themselves flag the bound-noun vs. RC boundary and a few predicate-class confusions (the seven jointly rejected cases). Collocation lists come from Lim (2009); that is fine if documented, but it is a fixed external list. None of this reverses the majority non-RC finding or the main claim. Citations look appropriate; no load-bearing math to break.\n\nThis is for people who work on Korean treebanks, UD mapping, or corpus syntax of adnominals. It will not reorganize general relativization theory, and it does not claim to. I would cite the distribution numbers and the released layer if I were annotating or evaluating Korean modifiers. A serious editor should send it to referees; the contribution is transparent and useful even if a referee wants a larger validation set. Engage with it.","headline":"Solid, reproducible corpus engineering: ETM is shared morphology, RC-like uses are only 39.4% of KLUE, and the annotation layer is released.","tokens_in":21494,"tokens_out":482,"would_cite":true,"duration_ms":4684,"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":"The Korean adnominal ending ETM marks several distinct noun-modifying constructions, not relative clauses alone.","keywords":["Korean","corpus annotation","adnominal constructions","relative clauses","treebank annotation","ETM","KLUE","construction typology"],"falsifier":"Re-annotate a large independent sample (or another Korean treebank) with the same typology and check whether relative-clause-like uses still fall near 39 percent, or whether morphology-plus-local-context rules recover a clear majority of true relative clauses that the current labels miss.","tokens_in":21591,"feed_emoji":"📝","tokens_out":913,"duration_ms":10227,"temperature":0.7,"pith_summary":"Korean uses one adnominal ending, ETM, across many ways of modifying nouns: relative-clause-like modifiers, adjectives, copulas, bound nouns, modals, temporal phrases, and fixed collocations. This paper shows that treating ETM as a relative-clause marker flattens real constructional differences and misleads corpus work. The authors build a typology based on predicate type, auxiliaries, argument structure, head-noun restrictions, and lexicalized patterns, then turn it into an ordered rule-based annotation layer on the KLUE dependency treebank. In the training data, productive relative-clause-like uses are only 39.4 percent of ETM instances; the rest are mainly adjectival, copular, bound-nominal, modal, temporal, and collocational. The practical point is clear: you cannot identify Korean relative-clause-like modification from morphology alone.","feed_headline":"Only 39% of Korean ETM forms are relative clauses","feed_subtitle":"The same ending also marks adjectives, copulas, bound nouns, modals and collocations—so morphology alone misleads.","key_machinery":"A construction-sensitive typology of ETM uses, operationalized as an ordered rule-based decision procedure (Algorithm 1) that tests collocational, adjectival, copular, and restricted patterns before productive relative-clause-like classes, then releases the resulting annotation layer over KLUE.","core_discovery":"ETM is a shared morphological exponent of several adnominal constructions, not a direct marker of relative-clause structure. In 13,046 ETM instances from the KLUE training split, productive relative-clause-like uses total only 39.4 percent; the majority are non-relative or construction-specific patterns that coarse modifier labels would collapse.","pith_inferences":["Any automatic relative-clause extractor for Korean that keys only on ETM will need an equivalent construction filter before its precision numbers are trustworthy.","The same diagnostics (predicate type, head-noun restriction, lexicalization) could transfer to other agglutinative languages whose adnominal morphology is similarly multifunctional.","Annotation schemes that collapse constructional subtypes into one modifier label will systematically inflate relative-clause counts in Korean learner and psycholinguistic corpora.","Borderline cases noted in validation suggest a useful next experiment: measure inter-annotator disagreement specifically on light nouns such as geos, jeom, and yejeong."],"forward_implications":["Morphology-based searches for Korean relative clauses will overgenerate roughly 60 percent of ETM tokens if they treat every ETM as a relative-clause marker.","Sejong-style coarse labels such as vp_mod underdistinguish construction types that the new layer makes recoverable.","Korean UD-style mapping of ETM across relations still needs an explicit constructional layer before cross-linguistic comparison is reliable.","Released labels and code let other researchers rebuild the annotation layer from original KLUE data under the same license.","Future treebank work can test the same typology on additional Korean resources and refine the bound-noun versus relative-clause boundary."],"fun_headline_variants":["Only 39% of Korean ETM forms mark relative clauses","Korean ETM is shared morphology across adnominal types","ETM flags relatives in just 39% of analyzed cases","Most Korean ETM uses are non-relative constructions","Relative clauses are not marked by ETM alone"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The ordered rules correctly encode linguistic precedence among construction types and do not systematically misclassify the borderline cases—especially bound-noun versus relative-clause and predicate-class ambiguities—that the small validation sample may have missed.","fun_headline_variants_meta":{"raw":{"variants":["Only 39% of Korean ETM forms mark relative clauses","Korean ETM is shared morphology across adnominal types","ETM flags relatives in just 39% of analyzed cases","Most Korean ETM uses are non-relative constructions","Relative clauses are not marked by ETM alone"]},"model":"grok-4.5","effort":"low","cost_usd":0.00384,"raw_usage":{"total_tokens":1194,"prompt_tokens":735,"num_sources_used":0,"completion_tokens":81,"cost_in_usd_ticks":38400000,"prompt_tokens_details":{"text_tokens":735,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":378,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":735,"tokens_out":81,"duration_ms":3442,"temperature":1.0,"reasoning_tokens":378,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:42:41.392037+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-annotate a large independent sample (or another Korean treebank) with the same typology and check whether relative-clause-like uses still fall near 39 percent, or whether morphology-plus-local-context rules recover a clear majority of true relative clauses that the current labels miss.","supporting_citations":[],"review_version":1}