PhD defense Tom Calamai: Identifying and Evaluating Misleading Climate Communication with Natural Language Processing

Jury

• Paolo Papotti, Professor, Eurecom, France (Reviewer)
• Manfred Stede, Professor, Universität Potsdam, Germany (Reviewer)
• Rachel Bawden, Chargée de Recherches, Inria, France (Examiner)
• Elena Cabrio, Professor, Université Côte d’Azur / Inria / CNRS / I3S, France (Examiner)
• Fabian Suchanek, Professor, Télécom Paris, Institut Polytechnique de Paris, France (PhD Supervisor)
• Oana Balalau, Inria Starting Faculty Position, Inria, France (PhD Co-supervisor)
• Théo Le Guenedal, PhD, Amundi Technology, France (Invited Member)
• Sofia Sakout, PhD, Amundi Technology, France (Invited Member)

Abstract

Climate communication is becoming more abundant, but not necessarily more informative. This thesis investigates to what degree Natural Language Processing (NLP) can help distinguish genuine climate-related claims from misleading content.

We first investigate the climate NLP literature and existing datasets to assess what current tools can already do for structuring climate communication and detecting potentially misleading content. While state-of-the-art models often perform well, many datasets are noisy, overly simple, or ambiguous, which limits the reliability of performance comparisons.

This shifts the focus from model performance alone to an additional bottleneck in climate communication assessment: insufficient disclosure. We therefore study corporate transition plans to identify what information is missing to properly evaluate their credibility. We propose an NLP pipeline to diagnose whether disclosed plans are specific, quantified, scoped, and detailed enough to be evaluated.

Finally, climate communication is not only a matter of disclosure. It is also argumentative, as companies justify their claims and frame their actions. We therefore examine fallacies in transition-plan discourse through a benchmark and an evaluation framework designed to account for ambiguity and subjectivity.