{"id":"db183f36-be3d-4d66-b41f-469526a37bef","arxiv_id":"2607.15238","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CLR-transformed unigram+bigram profiles classified by Aitchison distance reach 84%/95.6%/100% accuracy on short/medium/long texts across six Latin-alphabet languages.","lead":"A language identifier that treats letter and bigram frequencies as compositional proportions, maps them with the centered log-ratio transform, and classifies by Aitchison distance achieves 84–100% accuracy on six languages without neural training. It is deterministic and linear-time, making it a lightweight alternative for low-resource and edge deployments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 5's baseline ranges are not measured on LDDS; the 'slightly exceeds' claim is untested. Run langid.py/FastText on LDDS to settle.","rationale":"The paper's mathematical core is sound: the CLR isometry (Theorem 3.1) is correctly stated, and the method is deterministic and linear-time. The load-bearing issue is empirical. The central claim is comparative, yet Table 5 compares CoDA's measured accuracy on LDDS to baseline ranges quoted from other studies that were not run on the same data. Without a same-corpus baseline evaluation, the 'matches and slightly exceeds' sentence is unsupported. The reader identified this as the weakest assumption and issued a CONDITIONAL verdict; my read agrees. The concrete test—running langid.py and FastText on LDDS—would settle it directly. The paper's reproducibility pledge (code and dataset released) makes this test feasible. This is an addressable concern, not a fatal flaw, so no verdict change is warranted.","tokens_in":11399,"tokens_out":5097,"duration_ms":44278,"concrete_test":"Run the default langid.py model and a FastText classifier (trained on the same six reference corpora used for profile construction) on the exact LDDS test set, using the same lowercasing/tokenization and the same length bins (<50, 50–150, >150). Compare per-bin accuracies to Table 5. If either baseline meets or exceeds CoDA's 84.0/95.6/100.0, the 'slightly exceeds' claim fails; if CoDA remains higher, the concern is resolved. Also report the number of texts per bin and compute a significance test (e.g., McNemar's test or bootstrap confidence intervals) to assess whether observed differences exceed sampling noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—'matches and slightly exceeds the robustness of heavily optimized standard benchmarks' (§5.3)—rests entirely on Table 5, which places CoDA's held-out accuracies on LDDS against 'consensus accuracy ranges' for Raw Euclidean, Manhattan, Cosine, FastText, and langid.py taken from prior studies. No baseline is run on LDDS. The quoted studies use different language sets, class counts, corpora, text-length definitions, and preprocessing, so the ranges are not necessarily comparable to the six-language, 260-text task. The <50-character bin is likely small (the paper does not report per-bin counts), so a handful of misclassifications can shift percentages by several points. The OPUS experiment (§5.7) reports large-scale CoDA numbers but no baselines, so it cannot substitute for a controlled comparison. Unless baselines are evaluated on the exact same test set with identical preprocessing and length bins, the comparative superiority claim is untested. This is a measurement-comparability gap, not a flaw in the CLR mathematics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a deterministic, training-free language identification method based on compositional data analysis. Character unigram and bigram frequency vectors are treated as compositions on the probability simplex; additive Laplace smoothing is applied, and the centered log-ratio (CLR) transformation maps them into the (D−1)-dimensional zero-sum subspace, where Euclidean distance equals Aitchison distance. Classification is performed by computing the weighted sum of CLR-space Euclidean distances to fixed reference profiles for each language, with a length-dependent diacritic bonus term. The method is evaluated on six Latin-alphabet languages (English, German, Turkish, Romanian, Hungarian, Dutch). On the authors' held-out LDDS set (260 texts), it reports 84.0% accuracy for texts under 50 characters, 95.6% for 50–150, and 100% for over 150. A larger OPUS experiment (~250,000 sentences) reports similar aggregate numbers. The paper claims the method 'matches and slightly exceeds' the robustness of FastText and langid.py while remaining linear-time and fully deterministic.","tokens_in":11715,"tokens_out":5942,"duration_ms":46711,"significance":"The mathematical core is sound: Theorem 3.1 correctly states the CLR isometry between the simplex with Aitchison geometry and the zero-sum hyperplane with Euclidean norm, and the pipeline is genuinely linear-time and deterministic. The reproducibility pledge—public code, fixed hyperparameters, external profile corpora—is a notable strength. If the empirical comparison were made on a common testbed with direct baseline runs, the paper would offer a useful, interpretable, low-resource alternative to neural and embedding-based LID for alphabetic languages. However, as it stands, the central comparative claim is not yet established, because the baseline numbers are borrowed from other experiments rather than measured on the same held-out set.","major_comments":[{"comment":"The central claim 'matches and slightly exceeds the robustness of heavily optimized standard benchmarks' rests on Table 5, where the CoDA row is measured on the authors' LDDS test set but the baseline rows (Raw Euclidean, Manhattan, Cosine, FastText/langid.py) are 'consensus accuracy ranges' quoted from other studies. Those studies use different languages, corpora, class counts, preprocessing, and length definitions, so the columns are not comparable. On a 260-text dataset, the 84.0% vs. ~80–82% difference for short texts is also within plausible sampling error. To support the claim, the authors should run the baselines on LDDS with identical preprocessing and report per-bin counts and confidence intervals.","section":"§5.3, Table 5"},{"comment":"The empirical reporting is not sufficiently granular. LDDS contains 260 texts, but Table 6 gives per-language accuracy with one-decimal precision (e.g., 79.5%, 82.5%) without stating how many texts fall into each length bin; a few misclassifications can shift these numbers by several points. The ROC/AUC values are described as 'approximate' with no computational details. The OPUS experiment in §5.7 reports no baselines and no per-bin counts, so it corroborates scaling behavior but cannot substitute for a controlled comparison. Please report exact counts and, at minimum, 95% confidence intervals for the headline accuracies.","section":"§5.4, §5.6, §5.7"},{"comment":"The claim that CLR geometry is responsible for the improvement would be strengthened by an internal ablation that keeps the feature set and classifier fixed and only replaces the CLR distance with the same Euclidean/Manhattan/Cosine distance on the smoothed simplex vectors, evaluated on LDDS. The current Table 5 compares against quoted ranges, not the authors' own features/preprocessing, so it does not isolate the geometric contribution.","section":"§5.2–5.3"}],"minor_comments":[{"comment":"The phrase 'where Euclidean distances correspond to Aitchison distances' is repeated verbatim many times; please use it once and refer back.","section":"Abstract, §3.3"},{"comment":"The diacritic match count m is not formally defined. Specify how diacritic characters are identified per language and how m is computed.","section":"Algorithm 1, line 5"},{"comment":"Column headers list only L<50 and L>150, omitting the 50–150 interval used elsewhere. Include all three bins for consistency.","section":"Tables 3 and 4"},{"comment":"References [9], [10], [17], and [25] do not appear to be cited in the text; either cite them or remove.","section":"References"},{"comment":"The ROC figure is described in §5.6 but no plot appears in the manuscript text; ensure the figure is included and that AUC values are tied to actual bin counts.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper applies a standard mathematical tool (CLR) to a new domain; novelty is acceptable for a venue interested in lightweight NLP. The main problem is empirical comparability, not the mathematical core. I would be willing to accept after the authors run direct baselines on LDDS and report counts/CIs. The LDDS being an author-created dataset is acceptable given the public repository, but external validation would help."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is applying Aitchison's CLR transform to character unigram+bigram frequencies for language identification, with a length-dependent diacritic adjustment on top. The math is standard CoDA and the paper states it correctly, including the isometry theorem. The pipeline is deterministic, linear-time, and the reproducibility pledge plus the disclosed hyperparameter tuning on a separate validation subset are real credits. This is a legitimate domain transfer, not a subfield reshuffling, and the authors do not oversell it as one.\n\nWhat the paper does well: a clear formal setup, a sensible ablation showing bigrams matter, and an honest limitations section covering code-switching, script coverage, and the short-sequence smoothing bias. The absolute accuracies on their own LDDS data are plausible for six Latin-alphabet languages, and the OPUS experiment at ~250k sequences is a useful sanity check even without baselines.\n\nWhere it gets soft: Table 5 is load-bearing and it compares CoDA's held-out LDDS numbers against \"consensus accuracy ranges\" quoted from other studies — FastText, langid.py, Manhattan, Cosine are never run on LDDS. Those quoted numbers come from different corpora, language sets, and length definitions, so the \"matches and slightly exceeds\" claim in §5.3 is not actually supported by a controlled comparison. LDDS is a 260-text self-created test set with no reported per-bin counts, so sampling noise is a genuine concern; the difference between 95.6% and a baseline quoted at 93–95% could easily vanish or reverse with a proper head-to-head. This is a measurement-comparability gap rather than a flaw in the CLR geometry itself, and the stress-test note is right: run the baselines on the same texts and we will know.\n\nThe central mathematical argument holds up. The soft spot is empirical, not logical, and it is addressable. I would not cite this for the comparative claim, but the method description and the convergence-with-length story are worth knowing for anyone building low-resource or edge LID tools.\n\nFor peer review: yes, this deserves a serious referee. It is a coherent, reproducible application of established mathematics to a mature task, and the right editor would ask the authors to rerun their baselines on the same test set with identical preprocessing, add confidence intervals or exact counts, and preferably evaluate on an independent, larger corpus. Desk rejection would be a waste; it just needs the empirical work tightened.","headline":"Clean, honest application of CLR to language ID — the geometry is right, but the headline comparison to FastText/langid.py is not actually measured on the same test set, so the main empirical claim is untested.","tokens_in":12156,"tokens_out":1584,"would_cite":false,"duration_ms":15455,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T50","68T10","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Treating letter and bigram frequencies as compositional data and mapping them with a centered log-ratio transform yields a training-free, linear-time language identifier that matches or slightly exceeds tuned benchmarks on six Latin-script","keywords":["language identification","compositional data analysis","centered log-ratio transform","log-ratio geometry","n-gram frequencies","linear-time algorithm","deterministic classifier","zero-sum subspace"],"falsifier":"Use the paper's held-out test set of 260 texts, run the standard baseline identifiers cited in the paper (or faithful reimplementations of them) on those exact texts, and compare their length-stratified accuracies to the reported 84/95.6/100. If any baseline meets or exceeds those numbers, the paper's 'matches and slightly exceeds' claim would be falsified.","tokens_in":11339,"feed_emoji":"🌐","tokens_out":6207,"duration_ms":47100,"temperature":0.7,"pith_summary":"The paper argues that raw frequency vectors, because they sum to one, live on a simplex where ordinary Euclidean or Manhattan distances distort true relationships. It applies the centered log-ratio transform, which maps the simplex isometrically to a zero-sum Euclidean hyperplane, so that Euclidean distance becomes a valid compositional distance. On this geometry the authors build a classifier that compares CLR-transformed unigram and bigram profiles of a text against reference profiles, adds a length-decaying diacritic bonus, and predicts the closest language. On 260 held-out texts in six Latin-script languages, it reports 84.0% accuracy under 50 characters, 95.6% for 50–150 characters, and 100% above 150 characters, matching or slightly exceeding established off-the-shelf identifiers. The method requires no training, runs in linear time, and yields explicitly interpretable decisions.","feed_headline":"Log-ratio geometry gives 84-100% language ID without training","feed_subtitle":"By mapping letter frequencies into Euclidean space, it matches tuned baselines on six Latin-script languages.","key_machinery":"The centered log-ratio (CLR) transform: given a positive frequency vector x of dimension D, CLR(x)_i = ln(x_i) - (1/D) sum_j ln(x_j). This bijectively maps the D-simplex onto the (D-1)-dimensional zero-sum hyperplane, and the Euclidean norm on that hyperplane is exactly the compositional distance on the original simplex. The classifier uses this isometry twice: once for character unigram frequencies and once for bigram frequencies. The final score for a language is the sum of the two CLR-Euclidean distances to that language's reference profile, minus a term β·m/(1+α·L) that rewards diacritic matches on short texts.","core_discovery":"The central claim is that the centered log-ratio (CLR) transformation provides the right metric for language identification. For any positive frequency vector, CLR replaces each coordinate with its log minus the mean log, giving a vector that sums to zero; Euclidean distance in this zero-sum subspace equals the compositional distance. The paper constructs reference profiles for six languages from external corpora, then classifies a new text by summing the CLR-Euclidean distances of its unigram and bigram profiles to each reference profile, with a small length-dependent diacritic adjustment. On the held-out set it reports 84.0% accuracy for short texts, 95.6% for medium texts, and 100% for te","pith_inferences":["A testable extension beyond the paper's claims: run the same baselines on the identical held-out set to confirm whether the 'matches and slightly exceeds' wording holds under controlled comparison.","The length-decaying diacritic term is an engineered heuristic; a more principled version might emerge from Bayesian treatments of compositional observations.","The method's reliance on a fixed feature dimension suggests it may naturally extend to other alphabetic scripts (Cyrillic, Greek) with an aligned character set, but not to scripts whose symbol inventory varies.","For code-switched texts, a sliding-window variant of this classifier could localize language boundaries, since the CLR geometry projects mixed texts to intermediate points."],"forward_implications":["If the reported accuracy is reproducible, language identification for Latin-script texts can be performed with no training, using only precomputed frequency profiles.","The linear-time complexity (a single text pass plus fixed-dimension distance operations) makes the method suitable for real-time and edge deployment.","Because the feature space is fixed, adding a new language requires only constructing its reference profile, not retraining a model.","The interpretable geometry allows decisions to be traced to the distance contributions of unigram vs. bigram features.","The perfect long-text accuracy on the tested languages suggests that compositional separability improves with sequence length, pointing toward theoretical guarantees for sufficiently long inputs."],"fun_headline_variants":["Training-free language ID hits 84-100% via log-ratio","Compositional trick: CLR maps text to Euclidean space for 84-100% ID","No training needed: log-ratio geometry nails language ID up to 100%","Linear-time language ID: CLR transform yields 84-100% without training"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The comparison anchoring the paper's claim of parity with benchmarks relies on accuracy ranges taken from other studies on different test sets, not on running those baselines on the same 260 texts.","fun_headline_variants_meta":{"raw":{"variants":["Training-free language ID hits 84-100% via log-ratio","Compositional trick: CLR maps text to Euclidean space for 84-100% ID","No training needed: log-ratio geometry nails language ID up to 100%","Linear-time language ID: CLR transform yields 84-100% without training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000778,"raw_usage":{"total_tokens":3257,"prompt_tokens":705,"completion_tokens":2552,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":2464}},"tokens_in":449,"tokens_out":2552,"duration_ms":13940,"temperature":1.0,"reasoning_tokens":2464,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T23:45:43.711297+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the paper's held-out test set of 260 texts, run the standard baseline identifiers cited in the paper (or faithful reimplementations of them) on those exact texts, and compare their length-stratified accuracies to the reported 84/95.6/100. If any baseline meets or exceeds those numbers, the paper's 'matches and slightly exceeds' claim would be falsified.","supporting_citations":[],"review_version":1}