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

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.

• The paper demonstrates that machine learning models, including MLP and GuwenBERT, can detect regional linguistic fingerprints in Tang poetry with up to 0.69 accuracy.
• The paper employs character n-gram TF-IDF and domain-specific imagery features to differentiate North/South and multi-class regional groupings in a corpus of 49,000 poems.
• The paper concludes that regional imagery and lexical choices reflect both local stylistic distinctiveness and the effects of imperial cultural homogenization.

Computational Detection of Regional Linguistic Fingerprints in Tang Poetry

Problem Formulation and Dataset Construction

The paper addresses the longstanding question in literary history regarding the existence and detectability of regional linguistic signatures in Tang-dynasty poetry. Utilizing the Complete Tang Poems corpus, comprising approximately 49,000 poems from over 2,200 poets, the authors construct a poet-level dataset for 357 poets with ≥5 surviving poems, stratified across ten administrative circuits (dao) and further aggregated into binary (South vs. North) and three-way macro regional groupings. Geographic origin labels are derived from the China Biographical Database (CBDB), and stratified cross-validation alongside class weighting is employed to mitigate class imbalance, notably the dominance of Jiangnan and central/northern circuits.

Feature Engineering and Model Architectures

Two principal feature families are engineered: (i) character n-gram TF-IDF vectors (1–2 grams, up to 8,000 features, sublinear frequency scaling) and (ii) interpretable domain-specific features capturing the relative frequencies of imagery (mountain, water, plant, fauna, celestial), seasonal markers, allusional density, and type–token ratio. These features are concatenated to yield a poet-level representation summarizing their entire corpus.

For classification, a suite of models is utilized: logistic regression, linear SVM, random forest, MLP, and a fine-tuned GuwenBERT transformer. Classical models leverage balanced class weights and stratified 5-fold cross-validation, reporting both accuracy and macro-F1 scores against a most-frequent baseline. When comparing GuwenBERT, both fragment-level naive and hierarchical frozen encoder-pooled representations are evaluated to ensure parity with classical approaches.

Quantitative Results and Regional Signal Analysis

South/North region classification achieves a maximum accuracy and macro-F1 of 0.69 with MLP, outperforming the majority baseline of 0.53. Increasing target granularity to three-way and ten-class circuit groupings yields lower but still significantly above-chance macro-F1 scores (0.43 and 0.18, respectively). Peripheries—especially Jiangnan—display high identifiability (recall 0.71), whereas circuits near the political centers (Chang’an, Luoyang) are heavily confounded, indicating court-driven linguistic homogenization. Imagery emerges as the dominant regional marker, with southern poets favoring landscape motifs and northern poets exhibiting palace/gongti idioms; tonal prosody carries a weak signal due to standardization across regions.

A Mantel permutation test quantifies the linguistic distance decay effect ($r = 0.40$, $p \approx 0.09$): linguistic divergence grows with geographic separation but is concentrated in the distinctiveness of the Jiangnan idiom rather than smooth geographic gradients. Temporal analysis reveals that regional divergence is not static, with South/North separability at chance during High Tang (0.50) and peak differentiation during Late Tang (0.68), quantitatively supporting theories of imperial homogenization and subsequent regional autonomy.

Model Interpretability and Literary Historical Implications

Model misclassifications offer biographical insight: Early Tang errors exclusively misidentify southern poets as northern, reflecting the prestige of northern court idiom and its homogenizing effect. These errors are historically coherent, reinforcing the temporal modulation of the regional signal.

In a direct comparison, hierarchical GuwenBERT matches classical models (accuracy 0.674), and combining it with TF-IDF does not yield improvements, demonstrating that character-level n-gram statistics already encapsulate the available regional signal in classical Chinese poetic corpora.

Practical and Theoretical Implications

This study evidences that geographic-linguistic analysis, traditionally focused on spoken dialects, is translatable to ancient literary corpora, producing interpretable, hypothesis-generating models rather than opaque predictors. The persistence and fluctuation of regional signals in Tang poetry inform debates on regional schools, highlighting imagery and lexical choice as primary carriers of local distinctiveness, subject to court-driven temporal change.

Practically, the methodology provides a robust framework for literary scholars to surface latent patterns and generate informed hypotheses for further close reading. Theoretically, results suggest that centralized political power actively modulates cultural linguistic convergence, and court prestige can overwrite regional idioms, only to re-emerge as central authority declines.

Limitations and Future Directions

The sample size is limited (357 poets; 242 for binary classification), geographic and biographical labels inherit noise, and within-era analyses are constrained by small cohorts. Tonal feature extraction is approximated due to lack of complete rime dictionaries, and the distance-decay effect, though suggestive, would benefit from finer geographic granularity and increased sample size. The model is sensitive to corpus size and represents poets with entire corpora, which may mask intra-poetic variation.

Future work aims to extend temporal analysis and construct stylistic similarity networks to test whether textual proximity aligns with geographic distance or is mediated by literary influence, probing the interplay between spatial and cultural vectors in classical Chinese literature.

Conclusion

The geographic origin of Tang poets leaves a computationally detectable trace in their verse, carried predominantly by imagery and lexical choice, modulated by geographic and temporal factors. Interpretable machine learning models facilitate productive hypothesis generation for literary history, suggesting a dynamic interplay between regional linguistic distinctiveness and political integration. The approach opens avenues for expanded literary corpus analyses and finer linguistic-geographic modeling, with implications for both historical scholarship and computational humanities.