Скачать или смотреть [ProLaLa24] Large Language Models for Executable Tax Code Generation

Large Language Models for Executable Tax Code Generation (Video, ProLaLa 2024)
Sam Estep, and Vasudev Vikram
(Carnegie Mellon University; Carnegie Mellon University)

Abstract: The domain-specific language (DSL) Catala provides a way of formalizing legislative texts in order to write correct-by-construction implementations and resolve potential ambiguities. Motivated by their success in code generation domains, we propose using LLMs to generate executable Catala code to answer legal questions about the tax code. While we find that documentation and various prompting strategies can improve the compilation rate of LLM-generated Catala programs, we observe many difficulties with reasoning about the tax code, whose meaning is highly contextual. In response, we design Dagger, a new DSL that implicitly handles context in a way more akin to the way context is used in natural language and legal documents. We hypothesize this will create opportunities to further explore and improve the translation of legal documents into executable code.


Presentation at the ProLaLa 2024 conference, October 20, 2024, https://2024.splashcon.org/home/prola...
Sponsored by ACM SIGPLAN,