{"id":"c216ef72-d622-4416-bff5-6f32f8527600","arxiv_id":"2512.22487","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Korean constituency trees should use eojeol—not morphemes—as terminals, with morphology in a separate layer; normalizations can make Sejong, Penn, and KAIST comparable on that backbone.","lead":"This paper argues that Korean syntactic trees should treat the orthographic spacing unit (eojeol) as the terminal, with morphological details kept in a separate annotation layer. It proposes normalizations under which the major Korean treebanks can be compared on a shared eojeol backbone.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The equivalence claim in §5.1 is asserted, not tested: debinarization/null-removal are not shown to be inert, and Fig. 5's NML reanalysis selects among attachments fn. 4 admits are unresolvable. A corpus-scale normalization check is needed.","rationale":"The reader's CONDITIONAL verdict is appropriate. The paper is honest about the conditional nature of its equivalence claim and explicitly limits its scope; it reads as a position/design statement rather than a completed empirical study. But the abstract's 'comparative analysis shows' and §5.1's assertion of normalization inertness are not backed by any corpus-scale demonstration. The concern is not an internal contradiction; it is a missing empirical link, and it is directly testable. My analysis sharpens the reader's weakest assumption by pointing to two concrete mechanisms—debinarization can collapse distinct attachment structures, and the NML reanalysis selects among attachment options that the paper itself says cannot be decided on structural grounds. Because the proposed test could still confirm the equivalence, the appropriate verdict remains CONDITIONAL, matching the reader's verdict.","tokens_in":17911,"tokens_out":5992,"duration_ms":66396,"concrete_test":"Implement the §5.1 normalizations on the complete Sejong and Penn Korean treebanks. For every sentence, compute the set of constituent spans over eojeol terminals in (a) the original tree, (b) the debinarized/null-free tree, and (c) the final normalized tree. Then, for any sentences with identical surface eojeol sequences across the two treebanks, compare the normalized span sets directly; otherwise compare the transformation's span-loss/span-gain statistics on each corpus. If debinarization causes two originally distinct binary trees to yield the same flat span set, or if normalized Sejong and Penn span sets diverge on shared sentences, the inertness assumption fails and the equivalence claim is unsupported; if span sets remain identical after normalization, the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (§5.1) is that Sejong and Penn Korean constituency treebanks are representationally equivalent over eojeol after (i) debinarizing Sejong, (ii) removing Penn's null elements, and (iii) aligning eojeol tokenization. The load-bearing part is the assertion that these steps 'do not alter the set of eojeol terminals or their linear order, nor do they introduce new syntactic relations.' That inertness is never demonstrated, and it is not structurally obvious. Debinarizing a recursive NP→NP NP chain can collapse distinct modifier-attachment sites into one flat NP; two non-equivalent binary trees can map to the same flat tree, so the transformation can erase syntactic distinctions rather than merely remove notation. Penn null removal deletes an argument position and may leave an S-COMP dominating VP with no subject, a configuration whose equivalence to Sejong's clausal analysis is assumed. The proposal then adds an NML layer (§5.3, Fig. 5) and re-attaches modifiers; Footnote 4 gives three compatible analyses and concedes they 'cannot be resolved on structural grounds alone.' Thus the 'shared backbone' is not uniquely determined by the source treebanks. No corpus-scale comparison is reported, so the abstract's 'comparative analysis shows' currently outruns the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that Korean constituency treebanks should use eojeol (orthographic spacing units) as the terminal layer, with morphological segmentation and fine-grained POS information stored in a separate, explicitly non-constituent annotation layer. It reviews the Sejong, Penn Korean, and KAIST treebanks, proposes a set of normalization steps (debinarizing Sejong, removing Penn nulls, aligning eojeol tokenization, and applying the SPMRL 2013 conversion to KAIST) under which these resources are claimed to become representationally equivalent over a shared eojeol-based constituency backbone. On this basis, it outlines a six-column annotation format and discusses implications for cross-treebank comparison, constituency–dependency conversion, and future Korean resource development.","tokens_in":18245,"tokens_out":4155,"duration_ms":44910,"significance":"If the central equivalence claim held, the paper would provide a useful common structural baseline for comparing and converting Korean constituency treebanks, and its separation of constituency structure from morphological annotation is a clear and potentially valuable design principle. The paper is commendably transparent about the conditional, methodological nature of the equivalence, and the proposed six-column format is concrete and actionable. However, the load-bearing assertion that the normalizations are syntactically inert is never demonstrated, and the paper's own footnote concedes that key structural choices cannot be resolved on structural grounds. The significance is therefore contingent on additional validation, either through formal proof or corpus-scale empirical checks, which are currently absent.","major_comments":[{"comment":"The assertion that the three normalization steps 'do not alter the set of eojeol terminals or their linear order, nor do they introduce new syntactic relations' is load-bearing but unsupported. Debinarizing recursive NP→NP NP chains can collapse distinct modifier-attachment sites into a single flat NP; two different binary trees can map to the same flat tree, thus erasing syntactic distinctions rather than merely removing notation. Similarly, removing Penn's null elements deletes argument positions and may leave an S-COMP dominating a subjectless VP, changing the structural analysis of clausal complements. The paper should provide a formal definition of 'syntactic relations' and a proof of inertness, or support the claim with corpus-wide comparisons of original vs. normalized trees (e.g., checking that debinarization preserves head-modifier relations and that null removal does not alter","section":"§5.1"},{"comment":"The introduction of the NML layer and the re-attachment of modifiers in Figure 5 constitutes a further analysis choice, not merely a normalization. Footnote 4 explicitly states that at least three alternative analyses are 'compatible with the surface word order' and 'cannot be resolved on structural grounds alone.' This means the claimed 'shared backbone' is not uniquely determined by the source treebanks; the choice of Figure 5 is an annotation decision layered on top of the normalization. The paper should either demonstrate that the NML analysis is derivable from the original Sejong/Penn structures, or clearly separate this reanalysis from the equivalence claim and relegate it to an optional enrichment.","section":"§5.3 and Footnote 4"},{"comment":"The abstract states that a 'comparative analysis shows' representational equivalence, but the paper presents only a single worked example (Figures 1–5). No corpus-scale evaluation, statistics, or agreement measurements are reported across the Sejong, Penn Korean, or SPMRL-converted KAIST treebanks. If the equivalence claim is intended as an empirical result, it requires a corpus-level normalization study; if it is only a methodological conditional, the abstract and discussion should be rephrased to avoid claiming an empirical comparative result that the paper does not contain.","section":"Abstract and §6.1"},{"comment":"The treatment of KAIST relies on a cited SPMRL 2013 conversion procedure, but no example of a converted KAIST tree is shown, no description is given of how eojeol boundaries are inferred from the original morpheme-level terminals, and there is no verification that the converted trees satisfy the proposed eojeol-terminal condition. Without such evidence, including KAIST in the equivalence claim is unsupported and should either be substantiated with examples and checks or explicitly deferred to future work.","section":"§5.1 (KAIST)"}],"minor_comments":[{"comment":"Typo: 'Earlier resourceEarlier resources' should read 'Earlier resources'.","section":"§5.3"},{"comment":"Inconsistent capitalization of 'KAIST' vs. 'Kaist' across §3.3 and elsewhere; pick one consistent spelling.","section":"Throughout"},{"comment":"The statement that 'surface forms are the sole terminals' is potentially confusing because UPOS nodes are described as preterminals. Please clarify that the terminal level consists of eojeol, with UPOS labels as an intermediate preterminal layer, so the bracketing notation remains unambiguous.","section":"§5.5"},{"comment":"The six-column format is described in the text but the example is rendered with inconsistent whitespace, making it hard to see the column boundaries. Consider presenting the example with explicit tab separators or a clear table layout.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"This is a position/representation proposal with a prominent empirical-sounding claim. The central equivalence assertion is unsupported by the evidence provided. The paper would be considerably strengthened by adding a corpus-scale normalization experiment (e.g., convert Sejong and Penn Korean trees, measure agreement on flat structures, and check attachment stability) or by explicitly reframing the contribution as a proposal plus a worked example. Given the transparency and the potential utility of the proposal, I believe major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a position paper, not an empirical study. The authors propose that Korean constituency should be built over eojeol terminals with morphology in a separate layer, and they lay out a six-column format plus an optional NML layer. That is useful and mostly well argued. The paper is transparent about its scope: it explicitly says the equivalence between Sejong and Penn is conditional and methodological, and it concedes in footnote 4 that alternative NP analyses cannot be resolved on structural grounds alone. That honesty counts.\n\nWhat's new: not the idea that eojeol are terminals — Sejong and Penn already do that. What's new is the explicit normalization recipe (debinarize Sejong, remove Penn nulls, align tokenization) and the attempt to define a shared backbone plus an annotation format that separates morphology. That's a reasonable contribution to Korean treebank design and to discussion of terminal granularity for morphologically rich languages. The paper is clearly written, and the arguments against morpheme terminals are sensible.\n\nSoft spots: the central claim that the normalizations 'do not introduce new syntactic relations' is asserted, not shown. Debinarization can collapse distinct attachment sites into a flat NP; removing nulls changes argument structure; and the NML reanalysis in Figure 5 picks one of three analyses the authors themselves admit are equally compatible with the surface string. So the 'shared backbone' is not uniquely determined by the source treebanks. The abstract says 'a comparative analysis shows' — but there is no corpus-scale comparison, only a single worked example. The Kaist part leans entirely on SPMRL 2013. If the paper were framed strictly as a proposal for a normalization scheme, not as a demonstrated equivalence, the argument would be safe. As it stands, the equivalence is a design choice, not a finding.\n\nWho this is for: people working on Korean treebanks, parsing, or cross-framework conversion; also anyone thinking about terminal units in morphologically rich languages. They will get a clear articulation of one representational stance and a concrete format to react to.\n\nMy recommendation: send it to review. The paper is honest, well-scoped, and the issues are fixable — reframe the claim as a proposal, or run the normalizations over the actual corpora and show tree-changing rates. With that done, it could be a solid resource paper.","headline":"A clear, honest design proposal for eojeol-based Korean constituency; the conditional equivalence claim is asserted rather than demonstrated, so treat it as a proposal, not a result.","tokens_in":18684,"tokens_out":2173,"would_cite":true,"duration_ms":22462,"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":"Korean constituency trees should take eojeol as terminals and keep morphology in a separate layer.","keywords":["Korean constituency parsing","eojeol","morpheme segmentation","treebank normalization","constituency–dependency conversion","annotation scheme","universal part-of-speech","treebank equivalence"],"falsifier":"Take any sentence annotated in both the strictly-binary treebank and the null-element treebank, apply the paper's normalizations, and compare the derived eojeol-bracketed trees; if the two derived trees differ in constituent boundaries or labels for the same eojeol yield, the equivalence fails. A single such case, for example a null subject whose deletion changes where a complement attaches, would refute the claim.","tokens_in":17815,"feed_emoji":"🌳","tokens_out":5960,"duration_ms":57178,"temperature":0.7,"pith_summary":"The paper argues that Korean syntax should be annotated over eojeol—the spacing unit that usually combines a stem with case or ending—rather than over individual morphemes, and that morphology should live in a separate, non-constituent layer. Its central comparative claim is that after three explicit normalization steps (debinarizing one treebank, removing null elements from another, and aligning tokenization), the two major eojeol-based Korean constituency treebanks encode the same class of constituency relations, and the third, morpheme-based treebank can be converted to that same backbone. If true, the three resources become comparable and convertible without reconstructing word boundaries, and future parsing can treat eojeol as the surface terminal. This matters because existing Korean dependency resources are eojeol-based, so a shared eojeol backbone would remove a major representational mismatch between constituency and dependency grammar.","feed_headline":"Korean syntax trees should use eojeol, not morphemes, as terminals","feed_subtitle":"Normalization can give the three major Korean treebanks one shared structural backbone, with morphology kept in a separate layer.","key_machinery":"The eojeol — the orthographic spacing unit in Korean, typically a lexical stem plus functional morphemes — is the central object. The argument makes it the sole constituency terminal and treats it as the minimal unit that participates in syntactic relations. Working as the common backbone, it allows the three treebanks to be normalized into one representational space, and it anchors the six-column format in which constituency bracketing and morphological analysis are aligned but non-interacting.","core_discovery":"The paper's load-bearing claim is representational equivalence under normalization: debinarizing the strictly binary treebank, deleting null elements from the treebank that annotates empty categories, and standardizing eojeol tokenization do not change the set of terminals, their order, or the syntactic relations among them. Under those assumptions the two major eojeol-based Korean constituency treebanks can be treated as encoding the same eojeol-level constituency structure; the third treebank, which segments functional morphemes as independent terminals, is converted to the same backbone by reattaching those morphemes to their host eojeol. The paper then proposes a six-column annotation fo","pith_inferences":["A testable extension: if the equivalence is right, a constituency parser trained on one normalized treebank should transfer to another with little loss; unexpected performance gaps would reveal where the normalization did change structure.","The paper leaves open which of the three compatible nominal-modifier analyses to adopt; a systematic annotation-consistency experiment would settle it and would determine whether the proposed modifier layer's attachment decisions are reproducible.","The separation of morphology from syntax implies that morphological error does not propagate into syntactic evaluation, which would change how Korean parsing errors are diagnosed; this is an implication the paper does not develop.","If the equivalence holds across all three treebanks, then many previously reported cross-treebank syntactic differences may be representational artifacts (binarization, nulls, segmentation); re-analyzing published Korean parsing results on the normalized backbone would test how much of the variance survives."],"forward_implications":["The three existing Korean constituency treebanks can be compared and merged on a single eojeol backbone, so cross-treebank evaluation no longer requires reconstructing word boundaries.","Constituency–dependency conversion in Korean becomes a direct mapping because both representations share eojeol as their terminal unit.","Future treebanks can be built in layers: modifying the morphological segmentation does not change the syntactic tree, and vice versa.","Constituency parsing can be framed as bracketing over surface eojeol plus universal POS preterminals, independent of morphological analysis.","A practical path to open resources is converting public eojeol-based dependency corpora into the proposed constituency format, because the existing treebanks are license-restricted."],"fun_headline_variants":["Eojeol, not morphemes, as Korean syntax terminals","One eojeol backbone for three Korean treebanks","Korean constituency trees: eojeol terminals, morphology separate","Normalizing Korean treebanks to a shared eojeol structure","Eojeol-based constituency representation for Korean treebanks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the three normalization steps—debinarization, deletion of null elements, and tokenization alignment—are syntactically inert, altering neither the eojeol sequence nor the dominance relations among terminals; if any step changes attachment or hierarchy, the claimed equivalence collapses.","fun_headline_variants_meta":{"raw":{"variants":["Eojeol, not morphemes, as Korean syntax terminals","One eojeol backbone for three Korean treebanks","Korean constituency trees: eojeol terminals, morphology separate","Normalizing Korean treebanks to a shared eojeol structure","Eojeol-based constituency representation for Korean treebanks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1215,"prompt_tokens":689,"completion_tokens":526,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":433,"completion_tokens_details":{"reasoning_tokens":438}},"tokens_in":433,"tokens_out":526,"duration_ms":4852,"temperature":1.0,"reasoning_tokens":438,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T13:50:02.229703+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any sentence annotated in both the strictly-binary treebank and the null-element treebank, apply the paper's normalizations, and compare the derived eojeol-bracketed trees; if the two derived trees differ in constituent boundaries or labels for the same eojeol yield, the equivalence fails. A single such case, for example a null subject whose deletion changes where a complement attaches, would refute the claim.","supporting_citations":[],"review_version":1}