Cutting Cakes with Large Language Models - Yuval Marton, PhD, Computational Linguist and Artificial Intelligence / Data Science (AI / DS) expert.

Sentence-level semantics is often represented in terms of events: roughly, who did what to who (when, where and how). Each event has a semantic frame, specifying the possible participants or roles in the event: doer-of-the-action (agent), person-or-thing-done-to (patient or theme), done-with (instrument), done-at (location), etc. We can easily tell if such a role is well filled: “cut a cake with a knife” is good; “cut a cake with the sky” is not. (Psycho)linguists would say ‘sky’ has a low thematic fit to the instrument role of ‘cut’ in the sentence. Can we train large language models (LLMs) or other neural network models to estimate thematic fit well? Moreover, can this knowledge emerge in the model without direct training labels? If so, where would this knowledge be stored? Can we elicit it with careful prompt engineering from off-the-shelf LLMs? We continue a line of work in the intersection of Computational Linguistics and Psycholinguistics, and set a new state of the art in thematic fit estimation.

Yuval Marton, PhD, is a Computational Linguist and Artificial Intelligence / Data Science (AI / DS) expert, with a track record in both the industry (including IBM, Microsoft, Morgan Stanley, Bloomberg, and now Genentech) and academia (affiliate professor at University of Washington; industry mentor at Columbia University, UCSC and UMass). Dr. Marton’s experience spans Generative AI, multi-agent / agentic AI, retrieval-augmented generation (RAG), lexical semantics, paraphrasing, semantic role labeling, thematic fit estimation, parsing, statistical machine translation, dialog systems (a.k.a. personal digital assistants or chatbots), and more. He is also interested in increasing awareness to various forms of bias in NLP / AI, and how to develop and use these disruptive technologies ethically, given their substantial potential social impact for good or bad. Dr. Marton is an active Area Editor in the Association for Computational Linguistics (ACL) Rolling Review, served as the Multilingual Work and Translation Area Chair at the CoNLL 2024 conference, and co-organized several NLP workshops in top-tier international conferences. He received his Ph.D. in Linguistics from University of Maryland, concentrating on computational linguistics, with a Neuroscience and Cognitive Science (NACS) Program Certificate. He also holds a Masters in Computer Science from Polytechnic University (now NYU Poly).