Research

Publications · ORCID · dblp · HAL · theses.fr

Affiliation: associate researcher, SIG team (IRIT), in collaboration with Olivier Teste.

Research areas

1. Ontology alignment and explainability

I align ontologies: I look for which concepts in two knowledge models refer to the same thing. I carry out this work with Olivier Teste, within the international OAEI campaign (Ontology Alignment Evaluation Initiative).

Together we developed TOMATO (TOolkit for MATching Ontologies), presented at the 2022, 2023 and 2024 campaigns. TOMATO aligns ontologies two at a time. It computes several similarity measures between entity names, then a machine learning model decides which correspondences to keep. It competed every year on the Conference track, and on the Anatomy track in 2024. That year, we made it less dependent on reference alignments, in particular by training it on alignments generated by language models.

Before TOMATO, during my postdoc, I worked on holistic matching: aligning several ontologies at once rather than two at a time. No evaluation set existed for that case, so we proposed relaxed reference alignments, built from cliques in a graph (EKAW 2018, IC 2019). The Holontology system addressed this holistic matching at OAEI 2018 (results).

My line of work is now explainability. The questions I am working on:

In 2024, we tried decision trees to make TOMATO's model readable. They generalised poorly from one ontology to another; this work continues.

2. Accessibility and disability

My PhD dealt with text entry for visually impaired users; accessibility has never left my work since, and it also concerns me personally. I try to design digital tools that are simple, robust and genuinely usable by everyone.

During my PhD, I designed the DUCK keyboard (MOBACC 2013, IHM 2015, ICCHP 2018) and two word selection techniques, SWIFT (ICCHP 2016) and LOVIE (UAHCI 2016).

In 2023, with the CDDS in Rodez, I also ran development projects around disability with my students at 3iL Ingénieurs (the project).

3. Generative AI and disability

I study the way generative AI systems handle disability: which guardrails they apply, to whom, and with what asymmetries; and how those design choices concretely affect the people involved.

This area relies on an experimental approach: paired prompts, annotated datasets, reproducible figures. See Gemini, ChatGPT, Claude and disability (HAL preprint).

4. Generative AI and engineering education

As a teacher in an engineering school, I am interested in how students and teachers actually use generative AI, in its limits, and in the responsibilities it entails in education. The texts and talks I draw from it are gathered below, under Science outreach.

Tools and settings

The tools I develop for my courses are presented on profoutils:

Other tools are still in beta and not published: minivote, a Java lab in the browser (CheerpJ), Studium and annotation datasets.

Classifying long texts (with Miklós Sebők and co-authors)

This work sits between computer science and political science. It connects with a question that runs through areas 1 and 3: what a language model is worth on a specific task, evaluated on real data.

With Miklós Sebők and his co-authors, we classified laws and bills, in five languages, according to the 21 topics of the Comparative Agendas Project (arXiv, 2025). These texts often run to several hundred pages, far beyond what most models accept as input. We compared XLM-RoBERTa, which we had also used in TOMATO in 2024, Longformer, GPT-3.5 and GPT-4. Longformer, although designed for long texts, brings no clear advantage, and the best open model does better than the GPT variants. I carried this collaboration in 2025, despite my operation in January.

Supervision

Since 2026, Inès Taktak has been preparing a thesis entitled Alignement des données hétérogènes avec IA génératives (aligning heterogeneous data with generative AI), at IRIT (SIG team). Olivier Teste directs it and I co-supervise it, in collaboration with Faiza Ghozzi (University of Sfax). First milestone: taking part in the OAEI 2026 campaign, Bio-ML track, with the MetaMatch system.

Science outreach

I also explain what these systems are to non-specialist audiences: students, pupils in preparatory classes, newspaper readers.

Background

2012: internship at LIA (EPFL)

At the LIA at EPFL, I studied online social interactions. Network structure alone characterised user behaviour poorly. The emotional content of messages supplied the missing information (technical report).

2013 to 2017: PhD at IRIT

I prepared my PhD in the ELIPSE team at IRIT, supervised by Mathieu Raynal (manuscript). It dealt with text entry on smartphones for visually impaired people. On a touch keyboard, a blind user does not know exactly where their finger lands, and finding each key with a screen reader is very slow.

The DUCK keyboard accepts that imprecision: the user types approximately, and the keyboard deduces the possible words from the sequence of positions (MOBACC 2013, IHM 2015). What remains is picking the right word from a list. LOVIE optimises that list for non-visual selection (UAHCI 2016). SWIFT handles short words, where deduction has little information to go on (ICCHP 2016). I prototyped each technique, then evaluated it through experiments. DUCK was then evaluated in daily use (ICCHP 2018).

2017 to 2019: postdocs at IRIT (SIG team)

After the PhD, I moved from human-computer interaction to knowledge engineering. In 2017, a first six-month postdoc, on the ModelAGE project, brought me into the SIG team. The project aimed at a simulation support environment for predicting age-related pathologies (CSIT 2017).

From 2017 to 2019, in the regional SmartOccitania project, I worked on holistic ontology matching with Imen Megdiche, Cássia Trojahn and Olivier Teste (EKAW 2018, Holontology, OAEI 2018). That is where I started working on ontology alignment, which I continue today with TOMATO.

Since 2020: 3iL Ingénieurs and IRIT

After a year teaching at EPITA, I joined 3iL Ingénieurs, in Rodez, in 2020. I remained an associate researcher in the SIG team. From 2022 to 2024, I presented TOMATO at the OAEI campaigns with Olivier Teste. In 2025, after my brain surgery, I stopped my research and teaching for a few months, then gradually resumed, returning full time in January 2026. I now co-supervise Inès Taktak's thesis (Supervision).