Talks and Presentations
2025
Polarity inversion operators in PLM
29th Conference on Computational Natural Language Learning (CoNLL)
Vienna, Austria | July 2025
David Kletz, Pascal Amsili, and Marie Candito
Abstract: We investigate whether pretrained language models encode negation compositionally by training multilayer perceptrons to transform contextual representations to simulate polarity inversion. Our results suggest the existence of regularities compatible with the hypothesis of a unique compositional operator.
Swushroomsia at SemEval-2025 Task 3: Probing LLMs’ Collective Intelligence for Multilingual Hallucination Detection
19th International Workshop on Semantic Evaluation (SemEval-2025)
Vienna, Austria | July 2025
Sandra Mitrović, Joseph Cornelius, David Kletz, Ljiljana Dolamic, and Fabio Rinaldi
Abstract: We present a system for multilingual hallucination detection that leverages the collective intelligence of multiple LLMs to identify factual inconsistencies in generated text.
Better Together: Towards Localizing Fact-Related Hallucinations using Open Small Language Models
CHOMPS Workshop
Mumbay, India | December 2025
David Kletz, Sandra Mitrović, Ljiljana Dolamic, and Fabio Rinaldi
Abstract: We propose a method for fine-grained hallucination detection at the sentence level using combinations of open-source small language models, demonstrating competitive performance with large proprietary models.
2024
The Self-Contained Italian Negation Test (SCIN)
10th Italian Conference on Computational Linguistics (CLiC-it 2024)
Pisa, Italy | December 2024
Viola Gullace, David Kletz, Thierry Poibeau, Alessandro Lenci, and Pascal Amsili
Abstract: We extend the Self-Contained Negation Test to Italian, evaluating how Italian pretrained language models handle semantic constraints imposed by negation.
2023
The Self-Contained Negation Test Set
6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP
Singapore | December 2023
David Kletz, Pascal Amsili, and Marie Candito
Abstract: We introduce a novel evaluation methodology based on minimal pairs to assess whether pretrained language models have acquired knowledge of semantic constraints imposed by negation. Our self-contained test reveals that only very large generative models (LLMs like GPT-4) achieve near-perfect performance.
Probing structural constraints of negation in Pretrained Language Models
24th Nordic Conference on Computational Linguistics (NoDaLiDa)
Tórshavn, Faroe Islands | May 2023
David Kletz, Marie Candito, and Pascal Amsili
Abstract: We investigate whether pretrained language models encode structural constraints imposed by negation, focusing on the licensing of negative polarity items. Using probing classifiers, we demonstrate that contextual embeddings encode information about structural zones defined by negation scope.
2022
A Methodology for Building a Diachronic Dataset of Semantic Shifts and its Application to QC-FR-Diac-V1.0, a Free Reference for French
13th Language Resources and Evaluation Conference (LREC)
Marseille, France | June 2022
David Kletz, Philippe Langlais, François Lareau, and Patrick Drouin
Abstract: We propose a methodology for constructing diachronic datasets to study semantic shifts over time. We apply this methodology to create QC-FR-Diac-V1.0, a reference dataset for French covering newspapers from 1850 to 2000.
Invited Seminars and Research Talks
2025
Negation encoding in PLMs: Syntactic constraints and consequences
IDSIA NLP Group Seminar
Dalle Molle Institute for Artificial Intelligence, Lugano, Switzerland
December 2024
2024
Probing structural constraints of negation in Pretrained Language Models
NLP Group Seminar
Saarland University, Saarbrücken, Germany
October 2024
Negation encoding in PLMs: Syntactic constraints and consequences
ColingLab Seminar
Scuola Normale Superiore, Pisa, Italy
April 2024
2023
Probing structural constraints of negation in Pretrained Language Models
Lattice NLP Group Seminar
Lattice Laboratory (CNRS, ENS-PSL, Sorbonne Nouvelle), Paris, France
October 2023
Probing structural constraints of negation in Pretrained Language Models
LLF-LI Seminar
LLF Laboratory (CNRS, Université Paris Cité), Paris, France
September 2023
2022
Une méthodologie pour créer un corpus diachronique de changements de sens et son application à FDIAC 1.0, une ressource de référence en français
Rali-OLST Joint Seminar
Université de Montréal, Montréal, Canada
February 2022 (Remote presentation)
2021
Méthodologies pour la détection de diachronies sémantiques et leurs impacts
Lattice NLP Group Seminar
Lattice Laboratory, Paris, France
October 2021
Thesis Defense
PhD Thesis Defense: Négation et Modèles de Langue Pré-entraînés
Université Sorbonne Nouvelle, Paris, France
February 4, 2025
Thesis Committee:
- President: Mathieu Constant (Professor, Université de Lorraine) - Reviewer
- Reviewer: Marie-Catherine de Marneffe (FNRS Research Associate, UCLouvain)
- Examiner: Alexis Nasr (Professor, Aix-Marseille Université)
- Examiner: Guillaume Wisniewski (Associate Professor, Université Paris Cité)
- Supervisor: Pascal Amsili (Professor, Université Sorbonne Nouvelle)
- Supervisor: Marie Candito (Associate Professor HDR, Université Paris Cité)
Last updated: October 2025