Historical texts: getting by with a little help from AI
Lunchtime Lecture
This talk explores that question through the case of Medieval French and its dependency treeban PROFITEROLE. We first go through the methods which used for annotation, from earlier taggers to parsers and the use of LLMs. We then show how such annotated data can be explored, by researchers and curious readers alike, and ask what AI can realistically offer: a shortcut into historical texts, or a new companion that still needs expert guidance?
Abstract: Historical texts are notoriously hard to access: reading them takes years of training in philology and historical grammar, along with knowing where to find reliable editions, dictionaries and reference works. Linguists have long relied on Natural Language Processing (NLP) tools (lemmatisers, part-of-speech taggers, syntactic parsers) to annotate and query large corpora of earlier language stages. With the rise of large language models (LLMs), a new question arises: what can these tools actually do with languages nobody speaks anymore? This talk explores that question through the case of Medieval French and its dependency treeban PROFITEROLE. We first go through the methods which used for annotation, from earlier taggers to parsers and the use of LLMs. We then show how such annotated data can be explored, by researchers and curious readers alike, and ask what AI can realistically offer: a shortcut into historical texts, or a new companion that still needs expert guidance?
Mathilde Regnault is a postdoctoral researcher at the Institute of Linguistics / Romance Languages (ILR) at the University of Stuttgart, where she works with Achim Stein. Her research centres on heterogeneous data, from Medieval French texts to spoken child language. After working extensively on historical languages and syntactic parsing, including within the PROFITEROLE project, she now also works with CHILDES data and on pragmatics, exploring what these very different kinds of language data have in common and how computational methods can handle both.
Forschungsinstitut Brenner-Archiv
Georg-Trakl-Turm, Josef-Hirn-Straße 5 / 10. Stock
6020 Innsbruck
Teilnahme
Eintritt frei
Veranstalter
Forschungszentrum Digital Humanities