TL;DR
ALMeD is a new, semantically rich "gold-standard" corpus designed to fill a major gap in historical NLP: the lack of technical and scientific data for Medieval French. By providing high-quality annotations for medical and surgical treatises, the authors are building the foundation for more accurate linguistic analysis and specialized language models like the upcoming "HeidelBERT."
What Problem Does This Paper Address?
Natural Language Processing (NLP) for historical languages is notoriously difficult. Medieval French—a broad category spanning the 9th to 15th centuries—presents a "perfect storm" of challenges: there is no standardized spelling, syntax is constantly evolving, and high-quality training data is scarce.
While projects like the Syntactic Reference Corpus of Medieval French (SRCMF) exist, they primarily focus on literary texts (belles lettres) or legal documents. This leaves scientific and technical literature—which uses highly specialized terminology and distinct communicative practices—largely overlooked. Without domain-specific data, modern NLP models struggle to accurately parse or "understand" the technical nuances of medieval medical knowledge.
Methods
The authors developed ALMeD (the ALMA Medieval French Gold-Standard) using the INCEpTION annotation platform. Their approach is uniquely "fine-grained," moving beyond simple part-of-speech tagging to include deep semantic layers.
Key methodological features include:
- Multi-Layered POS Tagging: Every token receives two tags: a UPOS (Universal Part-of-Speech) tag for cross-linguistic standardization and a XPOS (language-specific CATTEX) tag for granular morphosyntactic detail.
- Dual Lemmatization: Tokens are mapped to a standard lemma (based on the Dictionnaire étymologique de l’ancien français or DEAF) and a "main lemma" representing the broader word family. This helps track etymological relationships despite wild spelling variations.
- Semantic Disambiguation: Unlike many corpora that stop at the word level, ALMeD includes glossary and dictionary definitions in both Modern French and English to resolve ambiguities in context.
Datasets
ALMeD is a work-in-progress corpus consisting of four primary medieval medical texts:
- Grande Chirurgie (Gui de Chauliac): Middle French treatises on history, tumors, and anatomy (~44,000 tokens).
- Surgery of Albucasis: A Lorraine scripta translation from the mid-13th century (~130,000 tokens).
- Lettre d’Hippocrate à César: An Anglo-Norman scripta from the 13th century (~24,000 tokens).
The authors use the DEAF dictionary as their primary lexicographic reference, supplemented by the Anglo-Norman Dictionary (AND) and the Dictionnaire du Moyen Français (DMF).
Key Insights and Ideas
- The "Main Lemma" Advantage: By grouping words into families, the researchers can capture the conceptual development of the language, making the dataset useful for both NLP engineers and philologists.
- Context is King in Medicine: The paper highlights how a common word like chief can mean "head" in one sentence and "tip of a bone" in another. Without the semantic disambiguation layer provided in ALMeD, a model would likely fail to distinguish these medical nuances.
- Standardization via UPOS: By using Universal Dependencies, the authors ensure their data can be aligned with other medieval Romance languages (like Old Occitan or Old Italian) being studied in the broader ALMA project.
Key Results
As a "gold-standard" resource paper, the primary result is the creation and public release of the dataset itself.
- Public Availability: A first version of ALMeD is already available on Hugging Face.
- Richness of Data: Every token in the gold standard contains a minimum of four labels (UPOS, XPOS, lemma, and main lemma), providing a much higher information density than previous Medieval French corpora.
- Foundation for HeidelBERT: The authors reveal that ALMeD is currently being used to develop HeidelBERT, a dedicated language model for Medieval French.
Limitations and Future Work
The most significant limitation is the current scale; while the total tokens across the four texts are numerous, the fully validated "gold-standard" portion is still growing. The authors aim to reach upwards of 50,000 tokens of semantically disambiguated text.
Future work will focus on:
- Completing the annotation and validation of the remaining treatises.
- Integrating the corpus into the development and evaluation of new NLP tools within the ALMA project.
- Expanding the dataset to include more regional varieties of Medieval French to better handle the "lack of spelling normalization" that remains a hurdle for automated systems.
Summary updated