iiinova — where your digital self keeps evolving

Raphaël Troncy @rtroncy.bsky.social · 10/05/2026

Slides from my #LLMText2KG 2026 keynote at ESWC are online: “From Strings to Knowledge Graphs: What Text2KG Still Gets Wrong.” drive.google.com/file/d/16kMb... Try #PERSEUS for Text2KG experiments: GitHub: github.com/Lettria/pers... Docs: docs.perseus.lettria.net #KnowledgeGraphs #LLMs

Take away message: what does it mean for a generated KG to be correct?

030

Raphaël Troncy @rtroncy.bsky.social · 30/06/2025

Can We Trust the Judges? This is the question we asked in validating factuality evaluation methods via answer perturbation. Check out the results at the #EvalLLM2025 workshop at #TALN2025 Blog: giovannigatti.github.io/trutheval/ Watch: www.youtube.com/watch?v=f0XJ... Play: github.com/GiovanniGatt...

031

Raphaël Troncy @rtroncy.bsky.social · 28/06/2025

Thanks to the jury members for having evaluated this work: Elena Demidov, Enrico Motta, Serena Villata, @papotti.bsky.social / co supervised with Pasquale Lisena

010

Raphaël Troncy @rtroncy.bsky.social · 28/06/2025

The final contribution is a rule-based system that again makes use of fine-grained causal relationships between events for explaining fact-checking results. Evaluated on the #FEVEROUS and #AVeriTeC datasets.

100

Raphaël Troncy @rtroncy.bsky.social · 28/06/2025

A third contribution is a method that relies on the JoinGT system but incorporates fine-grained causal relationships between events from FARO for generating narratives and stories of events. We demonstrated that qualitatively, the generated summaries are more complete.

100

Raphaël Troncy @rtroncy.bsky.social · 28/06/2025

This dataset can then be used to fine-tune language models (RoBERTa and REBEL) to be more effective at detecting fine-grained causal relationships between events in texts. A demonstrator making use of these models is available at demo.kflow.eurecom.fr

100

Raphaël Troncy @rtroncy.bsky.social · 28/06/2025

A second major contribution is the largest dataset containing fine-grained annotations of causal relationships between events (enable, prevent, intend, directly cause). This dataset comes from common sense (ATOMIC), news corpus (CNC) and synthetic #LLM generated data.

100

Raphaël Troncy @rtroncy.bsky.social · 28/06/2025

Youssra also investigated how capable #LLMs are for reverse-engineering competency questions and deriving user stories out of formal ontologies. Results demonstrate that recent large models such as GPT4 are effective for general domain.

100

Raphaël Troncy @rtroncy.bsky.social · 28/06/2025

A first contribution is the Facts and Events Relationships Ontology (FARO) that harmonizes different models representing event relations. This is available at anr-kflow.github.io/faro/

110

Raphaël Troncy @rtroncy.bsky.social · 28/06/2025

This is #PhD defense day at EURECOM. Youssra Rebboud is defending her 3.5 years research work on causal event relationship modeling, extraction and applications derived from them, scholar.google.com/citations?us...

110

Raphaël Troncy @rtroncy.bsky.social · 14/02/2025

Last demo to conclude #WAICF2025. Try out CimpleKG to fight #misinformation: ✅ Daily updated data from 77 fact-checking organizations ✅ Links 217k+ misinformation-related documents in 26 lang from 36 countries ✅ 15M+ triples, 263k+ entities & 1M+ textual features

020

Raphaël Troncy @rtroncy.bsky.social · 14/02/2025

We have presented @WAICANNES Conso Explorer, a joint work with @edfofficiel.bsky.social #SEQUOIA R&D team, that developed the largest #Knowledge #Graph about #electricity consumption in France. Billions of #RDF triples coming from #enedis #rte #meteofrance #insee #ign

120

Raphaël Troncy @rtroncy.bsky.social · 13/02/2025

Opening of the #WAICF2025 by the French Minister on Digital and AI @clarachappaz.bsky.social ! Come visit the booth E24 where EURECOM has an outstanding line up of demos over the next 2 days!

020

Raphaël Troncy @rtroncy.bsky.social · 11/02/2025

What a day at the #AI #Summit! Thrilled to report that EURECOM won the AI Frugal Challenge organized by @hf.co Our task: detect misinformation about climate change in texts The catch: be accurate but be very energy savvy! More to come with #ClimateSense, a new EU project

140

Raphaël Troncy @rtroncy.bsky.social · 11/02/2025

At the #AISummit Business Day listening the French and German minister of economic affairs. Proud to represent EURECOM for demonstrating #EULER, our innovative LLM-RAG tuned for Socratic interactions that aims to improve how students learn our graduate courses! #ai #education

020

Raphaël Troncy @rtroncy.bsky.social · 16/12/2024

Meet EULER (EURECOM Language Model for Education and Research) and our new #Socratic LLM for AI-Driven Education, eurecom-blog.medium.com/eurecoms-soc... Model available @huggingface.bsky.social and @ollama.bsky.social #LLM #RAG #EduTech

053

Raphaël Troncy @rtroncy.bsky.social · 27/11/2024

Fantastic talk from Serena Villata at the #SophiaSummit who is presenting the state of the art in argument mining. Datasets and tools available such as DispuTOOL and USElecDeb60To20. Supervised #BERT based models are still better than LLMs in particular for argument reasoning

020

Raphaël Troncy @rtroncy.bsky.social · 21/11/2024

More importantly, Julien is presenting a unique evaluation grid for qualitatively evaluating the #ontology automatically built by LLMs. Main results: ontologies are often too flat and insufficiently specific. #semwebpro Check out www.lettria.com for a demo

000

Raphaël Troncy @rtroncy.bsky.social · 21/11/2024

Can #AI and #LLM be used to facilitate the knowledge engineering process? Julien Plu from #LettrIA is presenting the ontology toolkit that leverages LLM to generate competency questions, classes, properties and their axiomatization using solely documents and a description of the domain.

121

Raphaël Troncy @rtroncy.bsky.social · 17/11/2024

"CimpleKG: A Continuously Updated Knowledge Graph on Misinformation, Factors and Fact-Checks", won the Best Resource Paper award at #iswc2024. Check out github.com/CIMPLE-proje... for the resource

070