Colloque international Intelligence artificielle & Intertextualité en langues anciennes, du lundi 26 octobre au mardi 27 octobre 2026 à Lyon (amphithéatre de la MILC).
Organisation : Laurence Mellerin et Théotime de la Selle
The rise of data science and the paradigm shift it has brought about has greatly encouraged interdisciplinary projects in historical sciences, particularly in the analysis of textual sources. Data-driven algorithmic approaches not only enable the automated processing of large volumes of text, but also the creation of new investigative tools. In the field of computational linguistics, large language models are highly effective at analysing and generating modern language texts. However, those for ancient languages are comparatively under-resourced and face specific challenges, such as limited linguistic resources, significant lexical and grammati- cal variability, and a lack of reference datasets for training and evaluating models. In recent years, projects combining artificial intelligence and the analysis of ancient texts have developed and refined tools for HTR (handwritten text recognition), NER (named entity recognition), lemmatisation and morphological and syntactic analysis. This has led to the creation of models, corpora and reference datasets for these algorithmic tasks. However, the intertextuality ecosystem remains largely underdeveloped, covering the detection and characterisation of textual reuses, text alignment, and authorship attribution. While tools based on lexical or syntactic approaches have become the benchmark for detecting literal quotations (Passim, TRACER, etc.), paraphrases and allusions appear to require semantic approaches enabled by machine learning.
This conference brings together researchers specialising in intertextuality in ancient languages from various countries to address the following two themes:
Development of models dedicated to intertextuality
Methodological reflection on machine learning models can draw inspiration from generic artificial intelligence work, such as paraphrase mining. However, the detection of reuses and text alignment differ from these more general tasks. Furthermore, the editorial context in which these tools may be used requires consideration of specific requirements, such as defining the reusing segment, integrating a typology of reuses and characterising the content and context of a reuse. How should we define the algorithmic tasks (similarity measures, classification, etc.) and learning strategies for models dedicated to intertextuality in ancient languages?
Creation of datasets and evaluation repositories
Many digital resources can provide reuse references or alignments. However, to our knowledge, there are currently no substantial, ready-to-use datasets to train supervised models. To facilitate collaboration between projects, it would be advisable to adopt common standards for data structuring. Is the sentence segmentation used in generic datasets relevant for alignment and reuse detection tasks? If so, what level(s) of segmentation granularity should be adopted, taking into account the specificities of these tasks and potential editorial requirements? Furthermore, since datasets are generally used for both training and evaluating models, how can we ensure that they adequately represent the diversity of case studies in order to serve as relevant evaluation tools? The labelling of allusions, which is particularly critical in the evaluation of tools, raises the issue of annotator subjectivity.