Pith. sign in

REVIEW 3 major objections 1 minor 8 references

Character n-gram TF-IDF models predict Tang poets' northern or southern origins from verse at 0.69 accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Machine learning on character n-grams from Tang poems predicts poets' broad regional origins above chance and reveals geographic distance-decay and temporal changes in linguistic separability.

T0 review reviewed 2026-06-26 challenge →

load-bearing objection Standard n-gram TF-IDF on the Tang corpus beats baseline for north-south origin and shows plausible time and error patterns, but the CBDB circuit labels and poem attributions receive no validation. the 3 major comments →

arxiv 2606.24093 v1 pith:VLAFULZJ submitted 2026-06-23 cs.CL cs.AI

Predicting Poets' Origins from Verse: A Computational Analysis of Regional Linguistic Fingerprints in the Complete Tang Poems

Chi-Sheng Chen , Hung-Yun Liu This is my paper
classification cs.CL cs.AI
keywords Tang poetryregional linguisticstext classificationn-gram TF-IDFliterary historyChina Biographical DatabaseGuwenBERTdistance decay
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aggregates every poem attributed to each of 357 authors in the Complete Tang Poems and links them to administrative circuits of origin drawn from biographical records. It treats the task as multi-class classification of poets by region using character n-gram frequencies plus features for imagery, season, and allusion. Classical and neural classifiers exceed the 0.53 majority baseline on the binary south-north distinction and remain above chance for the ten-circuit task. Three further patterns appear: linguistic distance between circuits tracks geographic distance, separability is at chance during the High Tang but rises later, and the model's confident mistakes in the Early Tang consistently map southern poets onto northern labels. A pre-trained classical-Chinese transformer matches but does not surpass the n-gram baseline.

Core claim

Regional origin leaves a recoverable linguistic trace in Tang verse. When poems are aggregated at the poet level and represented by character n-gram TF-IDF together with domain features, models achieve 0.69 accuracy on south-north classification against a 0.53 baseline and exceed chance on finer circuit prediction. Linguistic distances grow with physical distance between circuits, the signal is absent in the High Tang yet strongest in the Late Tang, and early-period errors align with the prestige of the northern court idiom. The same regional information is captured by simple n-grams as by a hierarchical frozen-encoder transformer, with no gain from their combination.

What carries the argument

Poet-level aggregation of verse from the Complete Tang Poems represented by character n-gram TF-IDF vectors plus interpretable domain features, used as input to origin classifiers.

Load-bearing premise

The biographical database correctly assigns each poet to a single circuit of origin and the poems attributed to that poet faithfully record his or her own linguistic habits without later editing or misattribution.

What would settle it

Accuracy falls to the majority baseline when the same models are retrained on only those poets whose circuit assignment is confirmed by at least two independent historical sources independent of the biographical database.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Linguistic distance between circuits increases with geographic distance, shown by a Mantel correlation of 0.40.
  • North-south separability is at chance levels in the High Tang and reaches its maximum in the Late Tang.
  • Confident misclassifications in the Early Tang always place southern poets into the northern class, matching the historical dominance of the northern court idiom.
  • A classical-Chinese transformer encoder matches but does not exceed the performance of the character n-gram model, and their combination yields no further improvement.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same aggregation-plus-n-gram pipeline could be applied to later dynastic collections to measure whether regional signals persist or erode under different political conditions.
  • The observed time interaction supplies a quantitative test for the hypothesis that imperial centralization temporarily overrides local linguistic markers in elite writing.
  • The method offers a way to estimate how much of an individual poet's output reflects birthplace versus later residence or court influence by comparing within-poet and between-circuit distances.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 1 minor

Summary. The paper aggregates all poems attributed to 357 Tang poets in the Quan Tang Shi, links each to one of ten administrative circuits of origin via the China Biographical Database (CBDB), and frames origin prediction as multi-class classification. Using character n-gram TF-IDF plus interpretable domain features (imagery, season, allusion), classical and neural models achieve 0.69 accuracy on the binary South/North task (vs. 0.53 majority baseline) and above-chance circuit-level performance. Additional results include a Mantel correlation (r=0.40) between linguistic and geographic distance, a temporal interaction in which South/North separability is weakest in the High Tang and strongest in the Late Tang, and historically interpretable error patterns; a frozen GuwenBERT encoder matches but does not exceed TF-IDF performance.

Significance. If the CBDB linkages and poem attributions prove reliable and confounds are ruled out, the work supplies quantitative evidence for regional linguistic variation in Tang poetry, a distance-decay effect, and court-driven homogenization followed by divergence. The demonstration that simple, interpretable n-gram features suffice (and that a domain-specific transformer adds nothing) positions the approach as a practical hypothesis generator for literary history.

major comments (3)
  1. [Data and Methods] Data linkage section: the entire classification pipeline and the Mantel correlation rest on the unverified premises that CBDB correctly assigns each of the 357 poets to a Tang circuit and that the Quan Tang Shi poems are representative of each poet's own linguistic habits. No cross-checks against primary sources (Jiu Tang Shu, Xin Tang Shu) or sensitivity analyses for label noise are reported; if these premises fail systematically the reported accuracies, distance-decay result, and temporal patterns become uninterpretable.
  2. [Results] Results section (accuracy and Mantel claims): the 0.69 South/North accuracy and the Mantel r=0.40 (p≈0.09) are presented without error bars, without specification of train-test splits or cross-validation procedure, and without ablation of the domain features. These omissions prevent assessment of whether the improvement over the 0.53 baseline is statistically robust or driven by the added features.
  3. [Results] Temporal-interaction analysis: the claim that South/North separability interacts with period (chance in High Tang, strongest in Late Tang) does not control for potential confounds such as shifts in genre distribution or poet birth-year cohorts across the four Tang periods; without such controls the observed pattern cannot be attributed to regional linguistic fingerprints.
minor comments (1)
  1. [Results] The p-value of approximately 0.09 for the Mantel test over nine circuits is marginal; the text should discuss effect size, multiple-testing considerations, and whether the result survives leave-one-circuit-out checks.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough review and valuable feedback on our manuscript. We address each of the major comments below, indicating the revisions we will make to strengthen the paper.

read point-by-point responses

  1. Referee: [Data and Methods] Data linkage section: the entire classification pipeline and the Mantel correlation rest on the unverified premises that CBDB correctly assigns each of the 357 poets to a Tang circuit and that the Quan Tang Shi poems are representative of each poet's own linguistic habits. No cross-checks against primary sources (Jiu Tang Shu, Xin Tang Shu) or sensitivity analyses for label noise are reported; if these premises fail systematically the reported accuracies, distance-decay result, and temporal patterns become uninterpretable.

    Authors: We agree that the reliability of the CBDB poet-circuit assignments is a foundational assumption. Although CBDB is the standard resource in the field for Tang dynasty biographical information, we did not perform independent verification against the official histories. In the revised manuscript, we will expand the Data section to include a discussion of CBDB's methodology and potential limitations. We will also conduct and report a sensitivity analysis by simulating label noise at various levels (e.g., 5-20% random flips) and re-evaluating the classification accuracies and Mantel correlation to demonstrate the robustness of our findings. revision: yes

  2. Referee: [Results] Results section (accuracy and Mantel claims): the 0.69 South/North accuracy and the Mantel r=0.40 (p≈0.09) are presented without error bars, without specification of train-test splits or cross-validation procedure, and without ablation of the domain features. These omissions prevent assessment of whether the improvement over the 0.53 baseline is statistically robust or driven by the added features.

    Authors: The original manuscript employed 5-fold cross-validation but did not report the per-fold variability or explicit split details in the main text. We will revise the Results section to include error bars representing standard deviation across folds for all reported accuracies, provide a clear description of the cross-validation procedure, and add an ablation study that compares performance with and without the domain features (imagery, season, allusion). This will clarify the contribution of each component and the statistical reliability of the results. revision: yes

  3. Referee: [Results] Temporal-interaction analysis: the claim that South/North separability interacts with period (chance in High Tang, strongest in Late Tang) does not control for potential confounds such as shifts in genre distribution or poet birth-year cohorts across the four Tang periods; without such controls the observed pattern cannot be attributed to regional linguistic fingerprints.

    Authors: We acknowledge that the temporal patterns could be influenced by changes in genre usage or the composition of poet cohorts over time. In the revision, we will extend the temporal analysis to include controls for genre distribution by adding genre proportions as additional features or covariates in the models. We will also examine the effect of birth-year cohorts by reporting results stratified by period-specific poet demographics where data permits. While these additions will help isolate the regional signal, we note that fully ruling out all historical confounds may require further data and will be discussed as a limitation in the revised paper. revision: partial

Circularity Check

0 steps flagged

No circularity: standard supervised classification on independent text features and external labels

full rationale

The paper extracts character n-gram TF-IDF and domain features directly from the Quan Tang Shi text, then trains classifiers to predict poet origin labels taken from the external CBDB database. No equation defines a quantity in terms of the target labels themselves, no fitted parameter is relabeled as a prediction, and no load-bearing premise reduces to a self-citation. The reported accuracies, Mantel correlations, and temporal patterns are evaluated against held-out data using standard cross-validation; the derivation chain therefore remains self-contained and non-tautological.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the accuracy of external biographical linking and on the assumption that n-gram statistics are not dominated by non-regional factors such as genre or date; no free parameters are explicitly fitted in the abstract but the choice of n-gram order and feature weighting are implicit modeling decisions.

free parameters (1)
  • n-gram order and TF-IDF weighting scheme
    Standard but unspecified choices that directly affect the feature space used for classification.
axioms (2)
  • domain assumption CBDB circuit assignments are accurate and stable labels for the poets
    Used to create the ground-truth classes for the multi-class and binary prediction tasks.
  • domain assumption Poems in Quan Tang Shi are faithful to the original author's linguistic habits
    Required for any origin signal to be detectable in the aggregated corpus.

reviewed 2026-06-26 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting Poets' Origins from Verse: A Computational Analysis of Regional Linguistic Fingerprints in the Complete Tang Poems." pith.science (2026). https://pith.science/paper/VLAFULZJ

@misc{pith2026260624093,
  author       = {Pith},
  title        = {Pith review of: Predicting Poets' Origins from Verse: A Computational Analysis of Regional Linguistic Fingerprints in the Complete Tang Poems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLAFULZJ}},
  note         = {Machine review of arXiv:2606.24093}
}
Share X Bluesky LinkedIn Reddit HN
abstract

We ask whether the geographic origin of Tang-dynasty poets leaves a detectable linguistic trace in their work. Aggregating every poem attributed to each author in the Complete Tang Poems (Quan Tang Shi) and linking poets to their administrative circuit of origin via the China Biographical Database (CBDB), we build a poet-level corpus of 357 poets across the ten Tang circuits and frame origin prediction as multi-class classification. Using character $n$-gram TF-IDF together with interpretable domain features (imagery, season, and allusion), classical and neural models predict a poet's broad region (South vs.\ North) at $0.69$ accuracy, well above the $0.53$ majority baseline, and finer circuit-level origin above chance. Beyond classification, three findings emerge. (i) Linguistic distance between circuits grows with geographic distance (Mantel $r=0.40$, $p\approx0.09$ over nine circuits), evidence of a distance-decay effect in poetic language. (ii) The signal interacts with time: South/North separability is at chance in the High Tang and strongest in the Late Tang, consistent with court-driven homogenization at the empire's height followed by regional divergence. (iii) The model's confident errors are historically meaningful -- in the Early Tang, every misclassification is a southern poet read as northern, reflecting the prestige of the northern court idiom. We further show that, when given the whole corpus through a hierarchical frozen-encoder representation, a classical-Chinese transformer (GuwenBERT) only matches -- not beats -- simple TF-IDF, and that combining them adds nothing, indicating that character $n$-grams already capture the regional signal. Our results position interpretable machine learning as a hypothesis generator for literary history.

Figures

Figures reproduced from arXiv: 2606.24093 by Chi-Sheng Chen, Hung-Yun Liu.

Figure 1
Figure 1. Figure 1: Poets per circuit (red = South, blue = North; poets with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Model comparison on South/North (5-fold CV). Dashed line: chance. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Circuit-level confusion matrix (row-normalized) and per-circuit identifiability (recall). Jiangnan is [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Identifiability (recall) rises with distance from the capital (Chang’an): peripheries are more dis [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Left: imagery features, South vs. North. Right: most discriminative characters per side (the bar [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Linguistic distance between circuits grows with geographic distance. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Left: corpus-length control (stable ⇒ not a length artifact). Right: South/North separability by era [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Misclassification direction by era. Early Tang: all South [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Fair comparison under identical 5-fold CV: hierarchical frozen BERT ties TF-IDF; the hybrid does [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

8 extracted references

  1. [1]

    and Varoquaux, G

    Pedregosa, F. and Varoquaux, G. and Gramfort, A. and Michel, V. and Thirion, B. and Grisel, O. and Blondel, M. and Prettenhofer, P. and Weiss, R. and Dubourg, V. and Vanderplas, J. and Passos, A. and Cournapeau, D. and Brucher, M. and Perrot, M. and Duchesnay, E. , journal =. Scikit-learn: Machine Learning in

  2. [2]

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , booktitle =

  3. [3]

    Liu, Yinhan and Ott, Myle and Goyal, Naman and Du, Jingfei and Joshi, Mandar and Chen, Danqi and Levy, Omer and Lewis, Mike and Zettlemoyer, Luke and Stoyanov, Veselin , journal =

  4. [4]

    2024 , howpublished =

    China Biographical Database (CBDB) , author =. 2024 , howpublished =

    2024

  5. [5]

    1705 , publisher =

    Quan Tang Shi (Complete Tang Poems) , author =. 1705 , publisher =

  6. [6]

    Cancer Research , volume =

    The detection of disease clustering and a generalized regression approach , author =. Cancer Research , volume =

  7. [7]

    2009 , publisher =

    Linguistic variation and computational dialectometry , author =. 2009 , publisher =

    2009

  8. [8]

    Sturgeon, Donald , journal =. The

This paper was first reviewed by grok-4.3 on June 26, 2026.