Stephen Langton's Quaestiones Theologiae

• Stephen Langton's Quaestiones Theologiae is a layered collection of scholastic theological questions derived from lecture reportation and reworked manuscripts.
• The corpus reveals multiple stages of transmission, reflecting variations from raw note-taking to edited literary forms across overlapping manuscript traditions.
• Advanced stylometric methods, including POS 3-grams and pseudo-affix analysis, uncover editorial influences and textual stratification in this multi-layered work.

Stephen Langton’s Quaestiones Theologiae is a corpus of scholastic theological questions that, in current computational and philological analysis, is treated not as a fixed authorial book but as a layered textual assemblage derived from classroom reportationes, reworked into literary form, and transmitted through multiple overlapping manuscript compilations. The collection originates in Langton’s Parisian teaching, probably from the last decades of the 12th century up to 1206, and its surviving forms preserve traces of reportatores, anonymous editors or compilers, and possibly Langton himself in some later corrections (Maliszewski, 18 Aug 2025).

1. Corpus profile and historical setting

The collection belongs to the environment of early university scholasticism and is especially significant because it preserves evidence for the transformation of oral teaching into written quaestiones. The underlying practice is reportatio: notes taken from oral teaching and subsequently reworked into literary form. Although indirect evidence suggests that such production was already not uncommon in the early scholastic period, direct evidence for how reportationes were produced and transformed is very scarce. Langton’s corpus is therefore valuable because it appears to preserve multiple stages of that process (Maliszewski, 18 Aug 2025).

The corpus is large and unstable rather than finalized. It was never given a final shape, contains clear signs of attempted editorial work, and includes many questions transmitted in multiple substantially different versions. Some versions are extremely concise and are plausibly close to raw reportationes, whereas others are expanded and polished literary forms. Of the 173 quaestiones in the contemporary index, 119 are transmitted in two to five versions. When the quaestiones extra indicem are included, the corpus amounts to over 350 distinct texts. A recurrent misconception is therefore corrected at the outset: the Quaestiones Theologiae is not a single-stage, uniformly authorial compilation, but a dossier-like body of material whose instability is constitutive of its transmission history.

2. Manuscript witnesses and stemmatic organization

The collection is transmitted by eight major manuscript witnesses:

Siglum	Manuscript
A	Avranches, BM 230
B	Arras, BM 965
C	Cambridge, St John’s College Library, C.7
H and K	Chartres, BM 430
L	Oxford, Bodleian, Lyell 42
R	Vatican, Vat. lat. 4297
S	Paris, BnF lat. 16385
V	Paris, BnF lat. 14556

Within this tradition, the codex C is internally complex and is divided into Ca–Cf. The manuscript evidence also supports discernible subcollections and families, especially H/K, $\alpha$, and $\beta$, with an additional $\gamma$ relation discussed in the stemmatic background. The argument drawn from this configuration is not that the tradition descends from a single copied exemplar through regular transmission, but that it reflects parallel, partially overlapping compilations of Langton’s material (Maliszewski, 18 Aug 2025).

This manuscript structure is central to interpretation. The subcollections are treated as independent editorial selections and reorganizations of lectures rather than merely mechanical witnesses to a stable archetype. That characterization matters because the stylometric program developed for the corpus is explicitly designed to test whether the textual stratification visible in the manuscripts also leaves detectable stylistic signals.

3. Formation model and editorial layers

The proposed model of formation is multi-stage. It begins with oral teaching, proceeds to reportatio by one or more anonymous recorders, then to reworking or expansion into a more literary quaestio, then in some cases to possible ordinatio or authorial correction by Langton, and finally to transmission through multiple manuscript compilations with distinct editorial profiles. The collection is thus approached as a collaborative textual product rather than a purely authorial one (Maliszewski, 18 Aug 2025).

The question of editorial responsibility remains open. Some material may have been reviewed by Langton himself, especially in ms. C, but most quaestiones were almost certainly edited by other people, possibly unknown students or secretaries from Langton’s milieu, and probably after 1206. The number of editors and reportatores involved remains unknown. The key interpretive proposal is that stylistic heterogeneity within the corpus may correspond to different layers of production. This suggests that variation across the corpus is not merely noise introduced by copying, but may index distinct editorial agencies embedded in the work’s textual history.

This unresolved distribution of responsibility is one of the central controversies surrounding the collection. The issue is not whether Langton stands behind the teaching that generated the corpus, but how far his control extended over the surviving literary forms. The protocol frames that problem in quantitative terms: whether stylometric clustering can help distinguish authorial correction, editorial reworking, and reportatorial mediation.

4. Stylometric design and feature engineering

The stylometric analysis follows work on collaborative medieval authorship, especially Kestemont, Moens, Deploige (2013), De Gussem (2017), and above all the workflow of Camps, Clérice, and Pinche (2021). Three feature types are proposed. The first is most frequent words, treated as a standard bag-of-words feature set focused on high-frequency, function-like lemmas or forms. The second is POS 3-grams, that is, sequences of Part-of-Speech tags of length 3, intended to capture syntactic structure beyond lexical choice. The third is pseudo-affixes, described as character n-grams at the edges of words; for the word verbum, the examples given are \_ve, ver, bum$, and um_ (Maliszewski, 18 Aug 2025).

An exploratory analysis based on the 200 most frequent words reports that the 10 most frequent words in the corpus are est, et, non, quod, in, ergo, set, ad, quia, and hoc. A PCA of 3,000-word samples from Langton, Robert of Courson, and Thomas Aquinas shows clustering by text of origin. The stated significance of that result is methodological: most frequent words can capture a real authorial signal even in scholastic Latin, despite the genre’s formulaic character.

At the same time, the protocol emphasizes a basic limitation of simple word-frequency stylometry for this corpus. Word-frequency analysis usually needs 2,000–5,000 words per sample to be reliable, but the average Langton quaestio is about 1,400 words; some are as short as 166 words, while the longest reaches 7,385. Concatenating questions would produce larger samples but would also suppress the very intra-corpus variation that the study seeks to detect. For each feature type, the minimal statistically reliable sample length is therefore to be computed using the test of Moisl (2011), following Camps, Clérice, and Pinche. The protocol also proposes LatinPipe or UDPipe 2 for POS tagging, citing reported accuracy of over 99% on Latin, including scholastic corpora.

5. Stemmatically informed regrouping and exploratory results

Because individual quaestiones are often too short for reliable clustering, the corpus is reorganized into 10 disjoint classes based on manuscript transmission. The classes listed are $\beta$, $\gamma$ with Cc–Cf, $\beta + \gamma + C$, Cb, $\gamma + Cb$, $\gamma$ without C, $\beta + C$, Ca, $\gamma + Ca$, and H/K. This regrouping is presented as a compromise between preserving meaningful stemmatic structure and obtaining samples large enough for stylometric analysis (Maliszewski, 18 Aug 2025).

When PCA is applied to these grouped classes using the 200 most frequent words, most classes cluster near the corpus average, but two clear outliers appear: Ca and H/K. The stylometric result is correlated with manuscript evidence. The unique material in Ca is located in final folios copied by a different hand, and the material transmitted only by H/K is likewise concentrated in later folios and may also have been copied alia manu. The importance of this observation lies in the alignment between computational signal and palaeographic observation. The protocol treats that alignment as evidence that stylometry may detect editorial or scribal heterogeneity rather than merely global lexical drift.

The regrouping strategy also reframes the object of analysis. Instead of assuming that each quaestio should be treated as an isolated authorial unit, it treats transmission-defined classes as possible carriers of editorial profile. A plausible implication is that manuscript families in this corpus are not only vehicles of preservation but also loci of compositional intervention.

6. HTR pipeline, registered-report hypotheses, and broader significance

A major methodological contribution of the protocol is the planned comparison between manually prepared data and automatically extracted HTR data. The study will use machine-readable texts from the critical edition alongside HTR-extracted transcriptions from manuscript images. The practical motivation is to include unedited manuscripts and manuscript-specific versions, extend corpora beyond what is currently edited, and provide a reusable workflow for scholastic Latin studies. The methodological question is whether an automated pipeline performs comparably well for stylometric work and whether editorial texts function as a kind of denoising relative to raw HTR output while possibly leaving detectable stylistic residue (Maliszewski, 18 Aug 2025).

The HTR system proposed is TrOCR, described as combining a vision transformer for feature extraction with a BERT-type decoder for character generation. The protocol attributes several advantages to this choice: it works well with normalized transcription, handles abbreviated Latin manuscript material more conveniently than CNN-based HTR systems, and may impose some orthographic normalization that could help later lemmatization and feature extraction. Even imperfect output is expected to be useful for POS tagging, pseudo-affix extraction, and other downstream stylometric features. The predicted character error rate is around 2–3%, but the protocol explicitly states that CER alone is not the relevant metric; the decisive issue is whether the extracted stylometric features remain reliable.

For segmentation and alignment, the workflow uses Kraken’s blla model. Existing editorial transcriptions are used first; a small portion of the material, about 20 pages, is manually aligned; a provisional Kraken model is then trained; and the remaining pages are automatically aligned. Automatic transcription alignment is based on the PASSIM script for text reuse detection, implemented in eScriptorium 0.13.

The study formulates three principal possible outcomes. First, distinct clusters among longer quaestiones may correspond to different editors, and one possible cluster could reflect material directly corrected by Langton. Second, short and long versions of the same quaestio may cluster together; if they do, that would suggest either that the reportator elaborated the text himself or that the longer version preserves most of the original reportatio, whereas failure to cluster together would imply a stronger stylistic difference between reportationes and literary quaestiones. Third, if enough short texts can be included, clusters of reportationes may be detectable and may link to individual reportatores. These are hypotheses rather than completed findings, because the study is a registered report protocol.

The broader significance is twofold. For Langton studies, the project may help determine whether the collection is best understood as a unified authorial work, a series of editorial reworkings, or a multi-layered compilation with several anonymous contributors. For the study of scholastic Latin and medieval university writing more generally, it offers a template for analyzing collaborative, orally derived academic writing by combining stylometry with HTR. The larger claim is not that computational methods replace philology, but that they can expose anonymous and collective intellectual labor that traditional philology alone cannot readily isolate.