NLP Post-doc / Engineer for Information Mining in Historical Data (French 3rd Republic)

1000scholars; Paris; On-site; EUR 42,000 - 54,000; Full time

Job summary

Inria ALMAnaCH, Paris, recrute un post-doctorant/NLP engineer pour concevoir et entraîner des modèles aidant les historiens à construire des sous-corpus à partir de vastes textes historiques numérisés.

Vous mettrez en place un flux d’annotation, évaluerez des classificateurs variés et favoriserez des solutions légères et déployables via une interface web destinée à des historiens non spécialistes.

Qualifications

• Compétences en Python et NLP pour entraîner et évaluer des modèles.
• Connaissances en traitement du langage naturel et en classification de textes historiques.
• Bonne maîtrise du français et de l’anglais, capacité à travailler avec des historiens.

Responsibilities

• Concevoir et entraîner des modèles pour aider les historiens à construire des sous-ensembles de corpus.
• Mettre en place un workflow d’annotation et évaluer des classificateurs.
• Intégrer l'apprentissage actif et viser des modèles légers et faciles à réentraîner.
• Assurer la compatibilité web et scaler pour des centaines de millions de tokens.

Skills

Python

NLP Training

Français

Anglais

Gestion de projets

Job description

Context

This position is part of the ANR project DECIDON (Débats parlementaires et Espace médiatique (1870-1940) : comprendre la CIrculation du discours politique grâce à des méthodes à forte intensité de DONnées), which is a collaboration between multiple French team, led at Epita Paris by Marie Puren, and including the Inria ALMAnaCH (Inria Paris center) project team as a work package lead for natural language processing techniques. The objective of the project is to enable a better understanding of how the parliamentary debates are being set up in the Third Republic and how thematics can appear and disappear across the newspaper and the parliamentary debate proceedings. The ALMAnaCH team role specifically focus on:

• Development of an annotation interface to support the creation of thematic datasets across the corpus of parliamentary debates and, potentially, the press.
• Development of an easily deployable approach for robust information retrieval across the corpus, targeting topics whose vocabulary may have evolved over time and diverged from contemporary French.
• Evaluation of information retrieval methods for topic-based search on a historically situated corpus (including RAG).

The employee will be supervised by Thibault Clérice, permanent researcher at Inria, and will work in close collaboration with other members of the team and project, namely Florian Cafiero and Marie Puren (EPITA). They will work at Inria, within the ALMAnaCH project-team. Within the team, they will find researchers connected to the topic outside of the project itself, including Cecilia Graiff, a PhD student in the team, supervised by Benoît Sagot and Chloé Clavel, working on multilingual and cross-cultural automatic analysis of argumentation structures in political debates.

Participation to national meetings and national/international conference are to be expected.

Assignment

The post-doc / NLP engineer will design and train models to support historians in constructing focused sub-corpora from large, noisy, OCR'd historical text collections. The work involves:

• Setting up an annotation workflow (true/false positive labeling of keyword occurrences in context) in collaboration with historians;
• Training and evaluating small-scale classifiers — ranging lightweight models to larger pretrained models — capable of distinguishing relevant from irrelevant occurrences of ambiguous or polysemous terms in diachrony (e.g., distinguishing an \"anti-parliamentary\" use of \"réforme de l'État\" from a routine administrative reform);
• Integrating active learning so that model performance improves iteratively as historians annotate, minimizing labeling effort while maximizing corpus quality;
• Prioritizing model deployability: given the scale of the corpus (at least dozens of millions of tokens across noisy OCR output) and the need for the tool to run efficiently and reproducibly within a web interface used directly by non-specialist historians, the research should focus on enabling this on lightway models: fast at inference, and easy to retrain/update rather than relying on large LLM inference at scale;
• Benchmarking against and complementing RAG-based exploration (T5.5), providing a transparent, low-cost alternative for corpus-scoping that historians can audit and reproduce before moving to more exploratory or generative tasks.

The research component focuses on efficient, low-resource sequence classification for historical/noisy text — including handling class imbalance, domain-shift across a ~70-year corpus, and figurative/contextual language. Publishable outputs (methods, benchmarks, and the resulting tool) are expected as part of the project's open, FAIR-data pipeline.

Bibliography:

El Assadi, A., Muennighoff, N., & Lee, J. (2026). The embedder's dilemma: LLMs are better, but at what cost? arXiv. https://doi.org/10.48550/arXiv.2608.12875

Martin, L., Muller, B., Ortiz Suárez, P. J., Dupont, Y., Romary, L., Villemonte de la Clergerie, É., Seddah, D., & Sagot, B. (2020). CamemBERT: A tasty French language model. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.645

van Strien, D., Beelen, K., Coll Ardanuy, M., Hosseini, K., McGillivray, B., & Colavizza, G. (2020). Assessing the impact of OCR quality on downstream NLP tasks. In A. Rocha, L. Steels, & J. van den Herik (Eds.), Proceedings of the 12th International Conference on Agents and Artificial Intelligence: ICAART 2020 (Vol. 1, pp. 484–496). SCITEPRESS. https://doi.org/10.5220/0009169004840496

Luo, X., Shinnick, Z., Griesshaber, N., Wang, Y., Yu, J., Shi, F., Torr, P., & Lu, Y. (2026). Pretraining language models on historical text. arXiv. https://doi.org/10.48550/arXiv.2606.02991

Bian, D., Puren, M., & Cafiero, F. (2026). How to efficiently explore noisy historical data? Leveraging corpus pre-targeting to enhance graph-based RAG. In D. Alves, Y. Bizzoni, S. Degaetano-Ortlieb, A. Kazantseva, J. Pagel, & S. Szpakowicz (Eds.), Proceedings of the 10th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature 2026 (pp. 241–250). Association for Computational Linguistics. https://doi.org/10.18653/v1/2026.latechclfl-1.23

Clerice, T. (2024). Detecting sexual content at the sentence level in first millennium Latin texts. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 4772–4783). ELRA and ICCL. https://aclanthology.org/2024.lrec-main.427/

Main activities

• carrying out research on the topic outlined above, both in the development of new ideas, positioning with respect to related work and validation of the methodology via experiments and analysis
• Producing a solution that will be deployable for post-lexical information retrieval
• Interact with the project’s historians as well as the researchers from the work package on natural language processing
• the presentation of work both internally to colleagues and externally in the form of conference/journal/workshop papers
• interacting and exchanging with colleagues on NLP topics

Skills

Technical skills and level required :

• Python
• NLP Model Training and evaluation, beyond just LLM

Languages :

• French (Reading understanding to interact with the data)
• English

Relational skills :

• Good organizational skills.
• Good interpersonal skills.

Additional skills considered an asset:

• Knowledge of the French 3rd Republic or similar political systems.
• Interefest in information retrieval
• A strong interest in open science.