Séminaire - DIC-ISC-CRIA - 10 septembre 2026 par Nicolas GOULET et Léo BOISVERT

Value of the Information & Strength of the Argument

The value of the information lies in its contribution to the symbol grounding problem and to understanding the structure of human language. The argumentation is solid: they start from established work on dictionary graphs (Vincent-Lamarre et al., 2016) and extend it methodologically by incorporating word sense disambiguation, which is a realistic improvement. They systematically test multiple dictionaries and languages, which strengthens the generality of the claim. The reasoning is careful, with explicit caveats about the approximation of human language acquisition by dictionary structure. The speakers also connect their findings to contemporary AI limitations, making the work relevant beyond cognitive science. However, the seminar format limits detailed proof of every step, and some technical details are glossed over, but overall the argumentation is persuasive and well-structured.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high: the speakers cite relevant prior work (Navigli, Vincent-Lamarre et al.) and present their own methodology in a transparent manner. The use of graph theory is appropriate and clearly explained. The sources provided in the description are academic and credible, though the talk itself does not include a live bibliography. The title of the video, as given in the parameters, is more generic than the specific content but still matches the overall theme. The talk is presented in French, which does not affect the scientific validity. No sponsorship or advertising segment was present. The public comments are not available, so no analysis of audience reception is possible.

The title accurately captures the core idea that a small set of words, through recombination, can generate the entire lexicon, which is the central message of the seminar.

8/10

The presentation is given by doctoral researchers with relevant expertise, and the research is based on a solid theoretical framework and computational analysis. The methodology is clearly explained, and the results are presented with caution. However, the video is a seminar recording without peer-reviewed publication at the time of the talk, so a score of 8 reflects high credibility with some residual uncertainty.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction by the host, introducing the speakers and the theme of the seminar series.
• 10:00Nicolas Goulet begins the presentation, introducing the concept of symbol grounding and its importance.
• 25:00Explanation of how dictionaries can be represented as graphs, with definitions as edges.
• 40:00Discussion of the MinSet concept and previous results showing it is about 1% of content words.
• 55:00Introduction of the new methodology involving word sense disambiguation and multilingual dictionaries.
• 1:10:00Presentation of results showing that MinSet remains small even after disambiguation across languages.
• 1:25:00Discussion of implications for language learning, pedagogy, and large language models, followed by Q&A.

• Is word sense disambiguation dead in the LLM era? — Cited by the speakers as a reference on the state of word sense disambiguation and its relation to large language models.
• Do large language models understand word senses? — Referenced to address the question of whether LLMs truly capture word senses, relevant to the discussion on grounding.
• Approaching the Source of Symbol Grounding with Confluent Reductions of Abstract Meaning Representation Directed Graphs — The speakers' own preprint, detailing the graph reduction methodology used in the research.
• The latent structure of dictionaries — Prior work by Vincent-Lamarre et al. that established the dictionary graph approach and the existence of a minimal defining core.