Learning a Planning Domain Model From Natural Language Process Manuals

Abstract

Artificial intelligence planning techniques have been widely used in many applications. A big challenge is to automate a planning model, especially for planning applications based on natural language (NL) input. This requires the analysis and understanding of NL text and a general learning technique does not exist in real-world applications. In this article, we investigate an intelligent planning technique for natural disaster management, e.g. typhoon contingency plan generation, through natural language process manuals. A planning model is to optimise management operations when a disaster occurs in a short time. Instead of manually building the planning model, we aim to automate the planning model generation by extracting disaster management-related content through NL processing (NLP) techniques. The learning input comes from the published documents that describe the operational process of preventing potential loss in the typhoon management. We adopt a classical planning model, namely planning domain definition language (PDDL), in the typhoon contingency plan generation. We propose a novel framework of FPTCP, which stands for a Framework of Planning Typhoon Contingency Plan, for learning a domain model of PDDL from NL text. We adapt NLP techniques to construct a ternary template of sentences of NL inputs from which actions and their objects are extracted to build a domain model. We also develop a comprehensive suite of user interaction components and facilitate the involvement of users in order to improve the learned domain models. The user interaction is to remove semantic duplicates of NL objects such that the users can select model-generated actions and predicates to better fit the PDDL domain model. We detail the implementation steps of the proposed FPTCP and evaluate its performance on real-world typhoon datasets. In addition, we compare FPTCP with two state-of-the-art approaches in applications of narrative generation, and discuss its capability and limitations.

NaRuto: Automatically Acquiring Planning Models from Narrative Texts

“…Few attempts have been made to generate planning models from narrative text, and even fewer have done so in a fully automated, unsupervised way. We compare the action domain models generated by NaRuto to those generated by Hayton et al (2017)'s StoryFramer, their 2020 system (Hayton et al 2020) (abbreviated "H2020"), and the results reported by Huo et al (2020). Of those, only H2020 is a fully automated system.…”

“…They avoid much of the complexity of dealing with narrative texts. Hayton et al (2017) and Huo et al (2020) explored narrative texts, yet their action model generation largely depends on manual interventions. Hayton et al (2020) proposed an automated process, but their action models aim only to mirror the input narrative.…”

“…Recently, researchers have attempted to extract action models from narrative texts such as short stories and movie synopses (Hayton et al 2017;Huo et al 2020;Hayton et al 2020). However, the methods proposed so far either generate quite simple and highly specific action models, or rely on human effort to complement or correct automatic extraction.…”

AI and climate resilience governance

“…Most common application domains associated with “policy appraisal” category is “energy and power management.” Such studies mainly focus on the following ( Figure 5 ): Improving communication and information sharing : among these are studies that use natural language processing to either connect user input to relevant knowledge discovery channels in the post-disaster communications—via smartphone application, web-based systems, and smart home devices 46 —or automate planning models for disaster response and recovery activities. 47 Some studies also use machine learning approaches to collect, integrate, and analyze data in order to improve communication and decision-making among humanitarian relief actors 48 , 49 and timely access to critical information of resources and services. 50 Improving resource allocation models : this includes studies that develop a type of decision support system for emergency operations to support efficient and fair allocation of limited resources following a disaster, e.g., automation of planning models.…”

“…Improving communication and information sharing : among these are studies that use natural language processing to either connect user input to relevant knowledge discovery channels in the post-disaster communications—via smartphone application, web-based systems, and smart home devices 46 —or automate planning models for disaster response and recovery activities. 47 Some studies also use machine learning approaches to collect, integrate, and analyze data in order to improve communication and decision-making among humanitarian relief actors 48 , 49 and timely access to critical information of resources and services. 50 …”