Plans for Monday 7/23

[Mark Liberman]

We'll meet at 4:00pm Eastern Time, on Zoom 
https://nyu.zoom.us/j/98819790926

The general topic is how to apply LLMs in evaluating the application of various kinds of abstract constraints to various kinds of documents (or events, or situations) -- where the answers are on a spectrum from "let the LLM (learn to) do it all" to "let the LLM (learn to) (help) create inputs for a domain-specific calculus".

We'll take a look at two specific cases, starting with the kinds of regulatory constraints described in

Kanovich, M., Kirigin, T. B., Nigam, V., Scedrov, A., Talcott, C., & Perovic, R. (2017). A rewriting framework and logic for activities subject to regulations. Mathematical Structures in Computer Science, 27(3), 332-375.

Andre Scedrov and Carolyn Talcott will lead us through a discussion of that paper, which proposes a pre-LLM approach. 

Alec Sugar and others from LexCheck will explain their current approach to checking whether proposed legal contracts (of certain types) meet user-specified constraints, and perhaps give some examples of how current LLMs do or don't solve all or parts of such problems.

Overall, one of our goals might be to look for a productive (real or artificial) exploratory task domain.

  See you tomorrow,

   Mark Liberman

   https://www.ling.upenn.edu/~myl