EVENT

Event News

Talk on "Textual Linguistics for Large Language Models" by Prof. Cavazza

We are pleased to inform you about the upcoming seminar byProf. Cavazza titled:"Textual Linguistics for Large Language Models" Everyone interested is cordially invited to attend!

Title:

Textual Linguistics for Large Language Models

Abstract:

Large Language Models (LLM) are generally associated to empirical natural language processing methods, and have even received criticism from traditional computational linguistics for their lack of connection to mainstream linguistic theories. However, it is widely accepted that semantic embeddings bear relations to distributional semantics, and some recent work has attempted to redefine traditional semantic concepts in the context of LLM. In this talk, I report an experiment exploring the relevance of textual semantics to LLM. Textual semantics describes the relation between semantic occurrences within a given text, through concepts such as semantic isotopies or lexical chains. In these experiments 10,000 short stories from an existing corpus have been completed by five different LLM of various sizes, origins, and training types: the resulting full stories are analysed for their semantic properties using gpt-4o as a judge, to determine whether completion preserves textual semantics. A number of metrics are introduced to support the analysis both in generic terms and for comparing individual LLM to each other, and to reference analyses published on 'human' texts. These preliminary results suggest that LLM may capture textual properties beyond the simple coherence of narrative topics.

Speaker Bio:

Marc Cavazza is Professor of Artificial Intelligence at the University of Stirling (UK), and a Visiting Professor at NII. His current research explores cognitive and linguistic aspects of Large Language Models. He has been active in the field for 30 years, starting with a PhD in natural language processing. He has participated in several large European research projects and has published in most international AI venues, including IJCAI, AAAI, AAMAS, ICAPS, ICML, and NeurIPS, receiving awards art ICAPS and AAMAS. He has served on the programme committee of over 200 conferences, and most recently has been a reviewer for NeurIPS, AAAI, ECAI, ICML and ICLR. He holds a PhD and an MD from University Paris Cité.