---
title: Scribal Profiling in Medieval Manuscripts
url: https://www.emergentmind.com/topics/scribal-profiling
type: topic
---

# Scribal Profiling in Medieval Manuscripts

Searching arXiv for recent papers on scribal profiling and related computational paleography.
arxiv_search query: scribal profiling medieval manuscripts paleography scribe identification
Scribal profiling, also termed “scribal modelling,” is the computational analogue of paleographic profiling: instead of identifying an author, it seeks to cluster and distinguish individual scribes on the basis of the linguistic and graphetic fingerprints they leave in manuscripts [2509.00067]. In current research, the term covers several related operations: text-independent off-line scribe identification from manuscript images, feature-based profiling from diplomatic or graphematic transcriptions, segmentation of a codex into homogeneous scribal runs, detection of hand shifts, regional or scriptorial clustering, and the analysis of positional regularities that reveal scribal behavior [2312.04296][2207.07726][2210.16780][2404.13069]. The field is shaped by a recurrent difficulty: medieval witnesses are layered objects, often copies of copies, so scribal signal must be disentangled from authorial strata, codex-specific artifacts, layout constraints, and editorial normalization [1802.01429].

## 1. Scope, targets, and problem setting

Scribal profiling addresses problems at multiple scales. At the most local scale, it treats scribe identification as a classification problem over patches, lines, pages, glyphs, or grapheme sequences. In the Codex Claustroneoburgensis database, for example, the task is defined as a $K$-way classification problem with $K=7$, where patch-level Softmax outputs are averaged to line level and then aggregated to page level [2312.04296]. At a larger scale, profiling can detect boundaries between hands inside a single codex, as in rolling stylometry for Cambridge, Corpus Christi 49, or fuzzy clustering over page-level feature rows in a 445-page archival volume [2207.07726][2210.16780]. At the broadest scale, it can recover regional groupings or scriptometric profiles from corpora of witnesses, as in Old French chansons de geste or the Herne manuscript production units [1802.01429][2509.00067].

The object of inference is not uniform across studies. Some work seeks individual hand attribution in a closed set of known scribes; some aims at profiling scribal behavior under positional or spatial constraint; some tests codicological hypotheses about collaboration, exemplar use, or manuscript transmission. The Hadewijch study measures the “degree of scribal appropriation” when ms. B was copied off ms. A, while the Herne study revisits earlier paleographic attributions through abbreviation practices and brevigraph-based stylometry [2210.14061][2509.00067]. In the Voynich Manuscript, the target is not writer attribution in the conventional sense but a catalog of positionally biased tokens that offers “a new handle on ‘scribal profiling’ in the Manuscript” [2404.13069].

A central methodological premise is that scribal identity is distributed across multiple weak cues rather than a single decisive marker. Old French scriptometrics explicitly states that “no single linguistic phenomenon allows to distinguish between different scriptae, and only the combination of multiple characteristics is likely to be significant” [1802.01429]. The same premise recurs in image-based work, where uniform scripts such as Carolingian minuscule make it difficult for classification to focus on meaningful features unless codex-dependent artifacts are controlled [2312.04296].

## 2. Transcription, corpus construction, and documentary representation

A major branch of scribal profiling is founded on transcription design. For medieval Latin Bibles, a “diplo-graphetic” transcription is used to preserve three layers of scribal information: letter-form variation, abbreviations, and orthographic or spelling variation [2207.07726]. Lowercase letter-forms such as “long-s” versus “short-s,” “insular-d” versus “rotund-d,” and $u/v$, $i/j$ distinctions are encoded rather than normalized away; superscript signs and macrons are rendered with combining diacritical characters; and common medieval abbreviations are kept in abbreviated form rather than silently expanded [2207.07726]. The explicit rationale is that transcription schemes “must be designed with specific research questions and scholarly end use in mind” [2207.07726].

Other corpora implement related but not identical policies. In the Herne corpus, manual “hyper-diplomatic” graphematic transcriptions and HTR-based Transkribus output preserve $u/v$ and $i/j$ distinctions, merge long-s and rotunda-r under graphematic standardization per MUFI, and retain brevigraphs [2509.00067]. In the Hadewijch corpus, line-break restoration rejoins hyphenated splits, medieval graphemes such as macrons and superscripts remain in the text, and graphemic transcription retains abbreviation signs while collapsing some allographic distinctions to base Unicode characters [2210.14061]. These differences are not contradictions so much as task-specific encodings of what counts as a “scribe-specific paleographic habit.”

HTR is frequently the enabling layer for large-scale profiling. The medieval Bible study reports that its best “composite” Paris-Bible model attains CER $\simeq 3\%$ [2207.07726]. In the Hadewijch case, Transkribus HTR+ models trained separately on ms. A and ms. B reached CER $= 2.52\%$ and CER $= 2.81\%$ respectively [2210.14061]. The Herne project uses 1,331 folios of manual transcriptions to train an HTR model with CER $= 2.7\%$, then applies that model to 4,828 further folios [2509.00067]. These results support the use of automatic transcriptions for downstream stylometry, although they do not remove the need for corpus-specific preprocessing and paleographic judgment.

Image-based profiling depends on a parallel representational discipline. In the Claustroneoburgensis study, line-level samples are automatically segmented by Transkribus; RGB line images are converted to grayscale, background is masked with a fixed threshold $\tau = 135$, lines are resized while preserving aspect ratio, and fixed-width windows are slid across the line to produce normalized patches [2312.04296]. In the YOLO-based writer-identification study, column crops are manually standardized per bible, red pixels are converted to white, and Otsu’s thresholding is applied after grayscale conversion [2601.04834]. The common principle is that representation is not neutral: preprocessing determines whether the model attends to ink traces, layout, parchment tone, rubrication, or noise.

## 3. Feature spaces and descriptive variables

Textual and graphetic profiling often begins with count-based feature inventories. One formulation defines a predetermined set $F=\{f_1,f_2,\dots,f_n\}$ of countable scribal features, including counts of abbreviations, special letter-forms, and orthographic variants, and represents a text segment $T$ by the vector
$$
f(T) = \bigl(\text{count}_T(f_1), \text{count}_T(f_2), \dots, \text{count}_T(f_n)\bigr)^T
$$
[2207.07726]. The Hadewijch study instantiates this strategy with a bag-of-words model over the 150 Most Frequent Words and a bag-of-character 3-grams and 4-grams over the 1,000 most frequent character-level $n$-grams [2210.14061]. Old French scriptometrics uses relative frequencies of the most frequent word-forms, testing cut-offs between ca. 600 and 3,000 MFW, with proper names and hapax legomena excluded [1802.01429].

A more explicitly graphematic formulation appears in the Herne study. It defines character-level abbreviation density as
$$
D_{abbr}(s)=\frac{\text{Number of brevigraphs by scribe }s}{\text{Total number of graphemes by scribe }s},
$$
and also a word-level counterpart based on abbreviated words over total words [2509.00067]. The same study uses a TF–IDF normalized bag-of-characters representation restricted to character bigrams that include at least one brevigraph, with $N \approx 1{,}493$ segments and $M=100$ frequent bigram features after filtering [2509.00067]. This is a deliberate attempt to emphasize “truly scribal choices, not content.”

Feature engineering can also operate below the token. Rajan’s metric suite for scripts proposes descriptors grounded in handwriting production and handwriting analysis, including length, divergence, bounding-box area, length–breadth index, average curvature, compactness, openness, distinctivity via DTW, ascendancy, descendance, circularity, rectangularity, stroke counts, changeability, disfluency, entropy, angle-based metrics, pen-drag distance, landmark-point count, and Ramer–Douglas–Peucker point count [1501.01894]. These metrics were developed for quantitative and descriptive analysis of scripts and “indirectly also quantify the scribal features” [1501.01894]. A plausible implication is that such descriptors supply a bridge between paleographic terminology and general-purpose pattern-recognition pipelines.

Image-first pipelines define their own feature spaces. For the Dead Sea Scrolls, textural “Hinge” features, allographic “Fraglets,” and weighted “Adjoined” concatenations capture complementary aspects of handwriting style [2010.14476]. For unsupervised forensic attribution in a historical manuscript, ten spatial features are extracted from word- and line-level bounding boxes, including stroke width, corner angles, convex area, orientation, height, width, aspect ratio, BlobLoG, BlobDoG, and connected components [2210.16780]. The Voynich study defines intrinsic cohorts such as TOP, FIRST, LAST, and MIDDLE, and extrinsic cohorts such as BEFORE and AFTER a drawing intrusion, then compares empirical glyph-count and token-propensity distributions across these positions [2404.13069]. Here the “feature” is positional context itself.

## 4. Modeling paradigms

Classical stylometric modeling remains central. Rolling stylometry based on Burrows’ Delta is used to segment a single codex into homogeneous scribal runs, while PCA plus $k$-means or hierarchical clustering operates on TF–IDF or word-form frequencies for witness-level profiling and regional attribution [2207.07726][1802.01429]. The Hadewijch study adds contrastive feature scoring through Burrows’s Zeta and uses a Random Forest with 500 decision trees over aligned character $n$-gram frequencies to identify the most distinctive features separating manuscript A from manuscript B [2210.14061]. The Herne study combines PCA, UMAP, one-class SVM, Random Forests, and $k$-NN with cosine distance [2509.00067]. These pipelines are interpretable in the narrow sense that they expose loadings, impurity decreases, nearest neighbors, and over- or under-represented forms.

A second family of methods targets unsupervised boundary detection. In the historical scanned-document study, StandardScaling and linear PCA reduce 100 handcrafted features to two or three dimensions, after which fuzzy $c$-means minimizes
$$
J_m(U,V)=\sum_{i=1}^N \sum_{j=1}^C u_{ij}^m \lVert x_i-v_j\rVert^2
$$
with cluster quality measured by the Fuzzy Partition Coefficient [2210.16780]. In the Great Isaiah Scroll study, PCA embeddings, $k$-means, nearest-neighbor voting, Chi-square tests, and a sigmoid change-point fit locate a likely switch around column 27 without requiring a prior assumption of writer identity [2010.14476]. Such methods are designed for cases where labels are absent, unstable, or historically contested.

Deep learning has introduced several distinct paradigms rather than a single dominant model. For medieval Carolingian minuscule, nine CNNs adapted to single-channel input and 7-way output were compared under a cross-codex design, including AlexNet, VGG, ResNet-18, DenseNet-121, Inception-v3, SqueezeNet, MNASNet, DeepWriter, and HalfDeepWriter [2312.04296]. Patch-level outputs $f(x_i)\in\mathbb{R}^K$ are averaged into a line-level vector
$$
f_{\text{line}}=\frac{1}{N}\sum_{i=1}^N f(x_i),
$$
and page-level decisions aggregate the line vectors [2312.04296]. The training setup uses Adam, learning rate $1.0\times10^{-5}$, batch size 32, 10 epochs, cross-entropy loss, early stopping on validation $F_1$, and a reject option threshold tuned on validation [2312.04296].

Other neural formulations are explicitly factorized. The Linear B study assigns each scribal hand a shared vector embedding and each sign type a shared vector embedding, then learns them jointly through reconstructive and discriminative losses so that “what sign is this?” is disentangled from “which scribe wrote it?” [2108.04199]. By contrast, the YOLO-based medieval-writer study collapses detection and attribution into a single grapheme-centered pipeline: YOLOv5s6 detects the letter “a,” and the model’s confidence score is interpreted as a thresholdable proxy for scribal attribution rather than feeding a separate classifier [2601.04834]. This suggests that scribal profiling can be framed at the level of full-page layout, line patches, isolated glyphs, or individual letter instances, depending on the historical question and the annotation regime.

## 5. Empirical results across corpora

Cross-codex CNN profiling on Carolingian minuscule demonstrates the importance of representation and split design. On Test B at patch level, RGB yields mean $F_1=0.30$, grayscale mean $F_1=0.53$, and masked grayscale mean $F_1=0.60$ [2312.04296]. DenseNet-121 reaches $F_1=0.61$ with 503 minutes of training, while AlexNet reaches $F_1=0.60$ in 115 minutes and is identified as the best trade-off between $F_1$ and time [2312.04296]. With AlexNet on Test B, individual-class page-level $F_1$ reaches 1.00 for scribe A 20 and 0.98 for A 259, while B 20 and A 215 show page-level $F_1=0.00$; the page-level confusion matrix suggests potential label noise for these systematically confused hands [2312.04296]. A reject option based on the maximum line-level Softmax score further stabilizes decisions: rejecting 20% of the most uncertain lines at $\theta\approx0.40$ reduces residual error by more than 50% [2312.04296].

Feature-based textual profiling produces strong separability when transcription preserves scribally informative detail. In the Hadewijch corpus, Random Forest classification over 1,000 character $n$-gram features yields typical results with accuracy $>95\%$ and $F_1 \approx 0.96$, and the most discriminative features isolate pronoun forms, abbreviation patterns, $h$-insertion, and suffixal preferences distinguishing manuscript A from manuscript B [2210.14061]. In the Herne corpus, abbreviation densities distinguish Herne scribes from non-Herne hands, one-class SVM outlier detection on Vienna SN 65 reaches Recall $=1.00$, $k$-NN on the $y$ versus $y(?)$ problem achieves Accuracy $=1.00$, and Random Forest separability for SN 65 versus other segments of scribe $a$ yields AUC $>0.95$ [2509.00067]. In the medieval Bible study, HTR-created transcriptions enable profiling of “individual and regional scribal characteristics,” and the case studies include perfect recovery of three hands in CCCC 49 with rolling stylometry on windows of 1,000 words and four regional clusters among 24 Paris Bibles through TF–IDF plus PCA [2207.07726].

Unsupervised and semi-supervised image-first studies show that hand shifts can be detected even without dense manual labels. In the 19th-century manuscript analyzed with fuzzy clustering, PCA$(k=2)$ plus 2-center $c$-means yields FPC$_1=0.86$ around the Hand$_1\leftrightarrow$Hand$_2$ shift and FPC$_2=0.72$ around the Hand$_2\leftrightarrow$Hand$_3$ shift, with top-1 grouping accuracy approximately 100% [2210.16780]. In the Great Isaiah Scroll, multiple independent feature methods show a bifurcation between columns 1–27 and 28–54; a secondary analysis yields a switching point around column 27, with a distance-comparison statistic $t\approx3.35$, $p=0.002$, one-way ANOVA $F\approx60.8$, $p<10^{-3}$, and a logistic fit locating $x_0\approx27$ [2010.14476]. In Linear B, the full disentangled neural model improves 5-fold macro-$F_1$ for findplace classification to 0.292 over a vanilla autoencoder at 0.234 and a most-common baseline at 0.154, and reaches a QVEC score of 57.0 against hand-crafted features [2108.04199]. In the YOLO-based medieval-bible experiment, template matching extracts 6,105 instances of “a” from Avila whereas YOLO extracts 202,294, and thresholding the confidence score at 0.83 yields Accuracy $=92.64\%$ and $F$-score $=81.78\%$ for attribution [2601.04834].

The Voynich Manuscript study broadens the empirical notion of profiling from hand attribution to behavioral regularities. Using 7,660 unambiguous tokens in 1,223 lines, it shows that token distributions vary significantly not only at intrinsic positions such as paragraph and line boundaries but also at extrinsic positions defined by drawing intrusions [2404.13069]. TOP and FIRST tokens are approximately 10–12% longer than MIDDLE, LAST and BEFORE are shorter, and these effects are significant under $\chi^2$ tests and token-level binomial/Bayes criteria with $p\le 0.01$ and $\log B \ge 5$ [2404.13069]. The paper concludes that extrinsic positional effects provide a statistically validated path toward profiling scribal behavior rather than only classifying hands [2404.13069].

## 6. Methodological tensions, limitations, and future directions

A recurrent tension in scribal profiling is the risk of learning the wrong signal. Cross-codex work explicitly warns that standard intra-codex training can overfit parchment- or layout-specific features rather than handwriting style, which is why it separates training codices from unseen test codices [2312.04296]. Scriptometric work raises an analogous concern on the textual side: witnesses are layered objects, so clusters can partly reflect a shared textual version or authorial fingerprint rather than only the scribe’s scripta [1802.01429]. The field therefore repeatedly tries to move the representation toward scribal choices—masked grayscale images instead of RGB, brevigraph-inclusive character bigrams instead of lexical content, or positional cohorts tied to writing-surface constraints rather than semantics [2312.04296][2509.00067][2404.13069].

A second tension concerns annotation and preprocessing. Line segmentation noise from Transkribus can degrade performance in CNN pipelines; fixed-threshold masking may not generalize; expanded abbreviations bias variant spellings in favor of majority editorial forms; low-contrast strokes and bleed-through reduce connected-component reliability; and manual cropping of glyphs remains labor-intensive in ancient-script datasets [2312.04296][1802.01429][2210.16780][2108.04199]. Some studies also report structural limits: cluster number $C$ must be chosen in advance in fuzzy clustering; the YOLO study focuses on a single grapheme, “a”; and several systems operate only in closed-set mode, fixed to known scribes [2210.16780][2601.04834][2312.04296].

Future work is correspondingly diverse. The Claustroneoburgensis study identifies open-set identification, transfer to other scripts or time periods, learned binarization, and active-learning loops with paleographer feedback as immediate extensions [2312.04296]. The Herne study proposes extension to other medieval vernacular corpora and regional dialect studies wherever abbreviation practices and letter-form variation are informative [2509.00067]. The Voynich study recommends applying the same cohort-based $\chi^2$ plus binomial/Bayes framework to other sections, other scribes, and other medieval manuscripts with mixed text/illustration layouts [2404.13069]. The YOLO study suggests extending from one grapheme to multiple letters and abbreviations and incorporating layout, stroke order, ink thickness, and ligatures in a multi-modal framework [2601.04834]. Taken together, these directions indicate that scribal profiling is not a single method but a family of approaches that joins digital paleography, HTR, stylometry, statistical testing, and neural representation learning in order to make scribal evidence computationally tractable while keeping it interpretable in codicological terms.

Source: https://www.emergentmind.com/topics/scribal-profiling