Daejeon, July 27–31
This project develops an open-data, TEI-based infrastructure for identifying, encoding, and modelling negative emotion in late-medieval English devotional and hagiographic texts. Its central aim is to show how emotions such as fear, shame, sorrow, silence, despair, humility, and bodily suffering do not function merely as descriptive states, but as structured narrative events through which female sanctity becomes visible, contested, recognised, or authorised. The corpus includes The Book of Margery Kempe, Julian of Norwich’s Revelations of Divine Love, the Constance tradition, and selected virgin-martyr hagiographies.
The project responds to a gap between medieval emotion scholarship and computational analysis. Work on medieval affect has shown that emotions operate within historically specific communities, devotional practices, bodily disciplines, and forms of female piety (Rosenwein; Bynum; Lochrie; McNamer; Downes and Trigg). At the same time, computational approaches to literary emotion often rely on lexical extraction, topic modelling, plot arcs, or sentiment tools trained on modern language data (Mohammad and Turney; Kim and Klinger). These methods are useful, but they are not designed to capture medieval emotional categories as historically situated moral, theological, and narrative structures. No current open resource encodes medieval negative emotions as relational data or tracks how emotional episodes produce patterns of examination, recognition, textual mediation, or spiritual authority.
The categories used in this project are therefore not adapted from modern affective computing. They are derived from scholastic and pastoral theories of the passions, especially Aquinas’s account of emotion as a movement of the sensitive appetite shaped by perceived good or evil, bodily alteration, and moral judgement (Aquinas; King). The project defines seven categories of negative affect: Shame, Humility_Subjection, Sorrow_Weeping, Despair, Fear, Bodily_Suffering, and Silence. These categories correspond to recurring affective patterns in late-medieval representations of female sanctity: public exposure, voluntary abasement, tearful lament, abandonment, dread, somatic affliction, and the interruption of speech under emotional intensity. The term “negative” does not imply moral failure. It designates states of vulnerability, pressure, subjection, or distress through which saintly women become subject to judgement and through which agency can paradoxically emerge.
Annotation is carried out in INCEpTION, a platform designed for custom annotation layers, linked annotations, and machine-assisted annotation workflows (Klie et al.). The project uses a three-layer schema:
AFFECT_TOKEN — linguistic items, gestures, bodily symptoms, weeping, silence, and affective expressions corresponding to the seven negative emotion categories.
AGENCY_TOKEN — passages in which authority, action, or recognition is activated, including public speech, confession, clerical examination, divine reassurance, writing/dictation, voluntary endurance, command, and communal response.
RELATION_LAYER — links between affective spans and agency spans, marking whether the connection is a trigger, transformation, or causal chain.
This architecture avoids treating emotional language as isolated data. Instead, it records how emotional events become narratively consequential. A passage of fear may lead to prayer or consultation; shame may lead to supervision, interrogation, or public recognition; silence may invite divine explanation or clerical interpretation. The relation layer does not assume that every co-occurrence is causal. Relations are encoded only when the text explicitly marks a consequential link through sequence, syntax, narrative framing, or theological interpretation. This responds to the risk of over-reading causality: the system distinguishes simple proximity from narratively marked transformation.
The agency layer is especially important because late-medieval religious culture regulated the forms of action available to women. Devotional practice, bodily suffering, and receptivity to divine intervention fit established models of female piety, while public speech, textual production, command, and doctrinal explanation entered domains associated with clerical or masculine authority (Aston; Hudson). The annotation model therefore tracks how texts justify female action through humility, suffering, divine confirmation, or clerical mediation. Agency is not treated as a general claim of empowerment. It is encoded as a limited set of historically legible actions through which women speak, endure, direct, are recognised, or enter textual circulation.
After annotation, data are exported as WebAnno TSV files and processed with Python. Scripts using Pandas and regular expressions normalise category labels, manage medieval spelling variation, align spans with the base text, and prepare the data for TEI conversion. A custom XSLT pipeline produces stand-off TEI-P5 files, enabling affective spans, agency spans, and their relations to function as searchable, reusable units within a TEI-compliant corpus (TEI Consortium). This format supports long-term preservation, interoperability, and future reuse by other scholars.
The database will be deployed in eXist-db, where XQuery enables retrieval of complex emotional sequences, such as all instances of Shame linked to Recognition_Mediation in hagiographic texts, or Fear linked to Confession_Testimony in visionary writing. TXM supports concordance, collocation, and distributional analysis of affective vocabulary across the corpus (Heiden). NetworkX generates graphs in which affective and agency events appear as nodes and relational links as edges, allowing repeated emotion–agency structures to be visualised across texts. Together, these tools create an open environment for querying medieval emotion as structured, relational, and interpretable data.
Once the initial dataset has been established and manually validated, the project will develop a domain-specific classifier capable of identifying negative emotion in unannotated Middle English and Anglo-Norman texts. This stage builds on the annotated corpus rather than replacing close reading. INCEpTION’s recommender system will first be trained on validated examples to suggest possible affective and relational spans. These suggestions will remain subject to manual review, ensuring that automation supports rather than determines interpretation.
The labelled data will then be exported to Python for supervised machine learning using tools such as spaCy, scikit-learn, or a compact transformer model fine-tuned on medieval textual data. Features will include lemmatised affective vocabulary, spelling variants, syntactic patterns associated with silence or fear, contextual markers of clerical judgement, and relational metadata derived from the project’s annotation layers. The goal is not to impose modern sentiment categories onto medieval texts, but to test whether historically defined categories can produce more responsible and meaningful computational models for under-resourced historical languages.
This contributes to broader debates in Digital Humanities about small data, domain-specific AI, and the limits of general-purpose sentiment analysis. The project argues that AI models for literary history should be calibrated to the conceptual worlds of the texts they analyse. In this case, medieval emotional semantics provide the training structure, not modern polarity or emotion-archetype systems.
Initial annotation reveals stable and repeatable structures across the corpus. In Julian of Norwich, episodes of fear, uncertainty, and bodily vulnerability are repeatedly linked to theological clarification, divine reassurance, and textual preservation. Affect becomes the condition through which revelation is interpreted and preserved.
In The Book of Margery Kempe, shame, weeping, accusation, and public disturbance generate forms of supervision, interrogation, travel, and recognition. Margery’s tears and social exposure repeatedly draw householders, priests, bishops, pilgrims, and scribes into sustained acts of judgement or mediation. Negative affect therefore becomes a mechanism of public visibility.
In hagiographic texts, humiliation, bodily suffering, and exposure often precede conversion, public recognition, or narrative reframing. The saintly woman’s body becomes the site through which violence is transformed into proof, and suffering becomes legible as sanctity only when it is recognised by witnesses, narrators, or divine intervention.
Across genres, negative emotion functions as a recurrent narrative mechanism through which women are corrected, protected, authorised, or made visible. The database makes these mechanisms comparable at scale. It allows scholars to ask whether shame more often leads to silence, speech, clerical examination, or recognition; whether fear produces withdrawal or devotional action; and whether bodily suffering is framed differently in hagiography, mystical writing, and lay devotional narrative.
The project delivers five main contributions. First, it creates the first open-data infrastructure dedicated to medieval negative affect. Second, it formalises a historically grounded emotional taxonomy in machine-readable form. Third, it develops a multi-layer annotation model that captures not only emotional events but also their narrative consequences. Fourth, it provides a reproducible technical pipeline from INCEpTION to Python, TEI, eXist-db, TXM, and NetworkX. Fifth, it establishes a roadmap for training AI systems on medieval emotional categories rather than modern sentiment labels.
The project also contributes to diversity in Digital Humanities by challenging the dominance of modern, English-language, social-media-based sentiment models. Its workflow is designed to be transferable to other historically specific emotional systems, including non-European traditions in which emotion is governed by distinct aesthetic, religious, or philosophical categories. Future collaboration could adapt the model to corpora structured by other emotional taxonomies, such as Sanskrit rasa theory, Arabic adab traditions, or religious literatures in other under-resourced languages. The broader aim is to offer a reusable method for building culturally specific emotion models rather than universalising modern affective categories.
This project demonstrates that medieval negative emotions can be rendered as structured, relational, and computationally searchable data without reducing them to modern sentiment. By combining philological precision, medieval moral psychology, TEI encoding, and machine-assisted annotation, it creates an infrastructure for analysing how affect shaped devotional authority, social interaction, and textual transmission in the late Middle Ages.
The resulting database supports both close reading and large-scale comparison. It allows scholars to move from individual scenes of fear, shame, sorrow, or silence to broader patterns of recognition, supervision, divine response, and narrative authority. More broadly, the project argues for a model of responsible computational humanities in which AI is trained on historically situated categories. Rather than asking medieval texts to fit modern sentiment systems, it builds computational methods from the emotional structures of the texts themselves.