Medieval Poetesses:
Emotional Intensity
in Historical Verse

01 · The Corpus

Four Voices, Four Worlds

The initial corpus centers on four medieval women whose textual traces are richly documented — chosen to maximize linguistic, geographic, and generic diversity.

Byzantine Empire · c.1083–1153

Anna Komnene

Political History

Author of the Alexiad — the first memoir written by a woman in the Western tradition.

Holy Roman Empire · 1098–1179

Hildegard von Bingen

Cosmology & Spirituality

Mystic, composer, scientist — her O Clarissima Mater maps the divine through language.

Anglo-Norman England · c.1160–1215

Marie de France

Courtly Love & Lais

Her Lais use natural metaphor to encode desire, longing, and social transgression.

Al-Andalus · c.1010–1091

Wallada bint al-Mustakfi

Independence & Passion

Poet and intellectual who hosted Córdoba's literary salon — and embroidered her verses on her robes.

02 · Data Collection

Automatic & Participatory

The corpus is built in two layers: automated scraping from open encyclopedic sources, and a participatory web app that invites researchers to contribute texts by medieval women from underrepresented regions.

Automated Scraping

A Python scraper queries the Wikipedia MediaWiki API targeting the categories Medieval_women_poets and Women_writers_(medieval). For each page, it collects author name, biographical extract, date, source name, and source URL.

BASE_URL = "https://en.wikipedia.org/w/api.php"
CATEGORIES = ["Medieval_women_poets", "Women_writers_(medieval)"]

params = {
    "action": "query",
    "list": "categorymembers",
    "cmtitle": f"Category:{category}",
    "format": "json"
}

The scraper is containerized with Docker and deployed on Quay.io, triggered automatically on each GitHub push.

Participatory Validation App

Scraped entries are reviewed and enriched through a FastAPI web app deployed on Render. Validated entries are written directly to Google Sheets, feeding the live corpus. The app is open to global contributors.

03 · Analysis Pipeline

From Raw Text to Emotional Map

Medieval texts in four languages are transformed into a comparable emotional signal through a four-stage pipeline.

04 · NLP — Linguistic Feature Extraction

Reading Emotion Across Languages

Medieval texts written in Old French, Latin, Byzantine Greek, and Andalusian Arabic are translated into English using a neural machine translation model before emotion classification.

• Neural Translation (NMT): Helsinki-NLP bridges medieval syntax and modern English so the emotion model receives data it was trained on.
• Emotion Classification: DistilRoBERTa calculates probability scores for 7 emotions (Joy, Sadness, Anger, Fear, Surprise, Disgust, Neutral) per line.
• Categorical Encoding: Emotion labels are converted to numerical codes for use as visual dimensions (color, size, position).

05 · Stylometry — Authorial Signatures

What Makes Each Voice Unique

TF-IDF vectorization weights words by how characteristic they are of a single author. PCA then compresses hundreds of linguistic dimensions into two axes — revealing the "gravitational pull" of each writer's style.

The connected scatterplot treats text as a sequence — linking verses in order to trace the tonal journey of a work, not just its average position in space.

• Sudden spikes represent thematic ruptures — moments of emotional or political intensity.
• Shared clusters between authors reveal cross-cultural structural resonance, not influence.

06 · Clustering — Thematic Connections

Patterns the Algorithm Finds Without Instruction

K-Means groups verses by semantic proximity without being told which author wrote them or what the theme is. Four clusters emerged as the optimal balance between granularity and interpretability.

07 · Timeline — Emotional Trajectories

Emotion as a Time Series

Narrative position functions as a proxy for emotional tempo. Each verse is treated as a moment in a diachronic sequence, allowing comparison of rhythmic structures across genres — historiography, theology, poetry — and across centuries.

The trajectory reveals not just what emotion dominates, but when a writer accelerates, plateaus, or ruptures into feeling.

08 · Historical GIS — Geography of Emotion

Where Feelings Were Written

Semantic clusters are projected onto the geographic nodes where these texts were produced. The map asks: does proximity in emotional space correlate with, or contradict, proximity in physical space?

Córdoba, Bingen, Caen, and Constantinople sit thousands of kilometers apart — yet Cluster 04 (Loss/Memory) draws Wallada and Anna Komnene into the same mathematical neighborhood.

The cartography maps four geographic nodes to their dominant emotional cluster:

09 · Key Findings

What the Algorithm Reveals

Rather than confirming identity-based assumptions, the analysis surfaces structural patterns that emerge from the text itself.

10 · Global Corpus & Contribution

Voices Still Waiting

The current corpus reflects what is digitally accessible in European archives. The project actively seeks texts by medieval women from Africa, Asia, and the pre-Columbian Americas.

11 · Citation & Methodology

How to Cite

Keywords: digital humanities · computational philology · emotion analysis · historical GIS · medieval women's writing · NLP · time series · unsupervised learning · K-Means · PCA · attentive reading

Corpus shared under Creative Commons CC-BY 4.0. Source code on GitHub ↗