References & Citations
Computer Science > Formal Languages and Automata Theory
Title: Automaton-Based Representations of Task Knowledge from Generative Language Models
(Submitted on 4 Dec 2022 (v1), last revised 9 Aug 2023 (this version, v5))
Abstract: Automaton-based representations of task knowledge play an important role in control and planning for sequential decision-making problems. However, obtaining the high-level task knowledge required to build such automata is often difficult. Meanwhile, large-scale generative language models (GLMs) can automatically generate relevant task knowledge. However, the textual outputs from GLMs cannot be formally verified or used for sequential decision-making. We propose a novel algorithm named GLM2FSA, which constructs a finite state automaton (FSA) encoding high-level task knowledge from a brief natural-language description of the task goal. GLM2FSA first sends queries to a GLM to extract task knowledge in textual form, and then it builds an FSA to represent this text-based knowledge. The proposed algorithm thus fills the gap between natural-language task descriptions and automaton-based representations, and the constructed FSA can be formally verified against user-defined specifications. We accordingly propose a method to iteratively refine the queries to the GLM based on the outcomes, e.g., counter-examples, from verification. We demonstrate GLM2FSA's ability to build and refine automaton-based representations of everyday tasks (e.g., crossing a road), and also of tasks that require highly-specialized knowledge (e.g., executing secure multi-party computation).
Submission history
From: Yunhao Yang [view email][v1] Sun, 4 Dec 2022 22:34:16 GMT (657kb,D)
[v2] Fri, 3 Feb 2023 19:27:35 GMT (1145kb,D)
[v3] Thu, 16 Mar 2023 15:27:46 GMT (1345kb,D)
[v4] Wed, 10 May 2023 20:10:33 GMT (1179kb,D)
[v5] Wed, 9 Aug 2023 21:55:36 GMT (1679kb,D)
Link back to: arXiv, form interface, contact.