NLtoPDDL
One-Shot Learning of PDDL Models from Natural Language Process Manuals
Shivam Miglani (Student TU Delft)
Neil Yorke-Smith (TU Delft - Electrical Engineering, Mathematics and Computer Science)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Existing automated domain acquisition approaches require large amounts of structured data in the form of plans or plan traces to converge. Further, automatically-generated domain models can be incomplete, error-prone, and hard to understand or modify. To mitigate these issues, we take advantage of readily-available natural language data: existing process manuals. We present a domain-authoring pipeline called NLtoPDDL, which takes as input a plan written in natural language and outputs a corresponding PDDL model. We employ a two-stage approach: stage one advances the state-of-the-art in action sequence extraction by utilizing transfer learning via pre-trained contextual language models (BERT and ELMo). Stage two employs an interactive modification of an object-centric algorithm which keeps human-in-the-loop to one-shot learn a PDDL model from the extracted plan. We show that NLtoPDDL is an effective and flexible domain-authoring tool by using it to learn five real-world planning domains of varying complexities and evaluating them for their completeness, soundness and quality.