Towards Comprehensive Automation: Extracting Action Models from Diverse Scenarios

Abstract

AI planning is dependent on the systematic formation of action sequences or plans that transition from a given initial state to a desired objective state. To accomplish this, well-defined action models, which are typically encapsulated in planning languages such as Planning Domain Definition Language (PDDL), are required. However, creating these models by hand requires considerable effort and domain-specific expertise.

This thesis focuses predominantly on the automated generation of action models from two crucial domains: physical domains and narrative texts. Fundamental to our methodology for the physical domains is the incorporation of qualitative spatial relations to represent action states. Regarding the narrative texts, we present NaRuto, an automated system that generates planning action models efficiently from narrative texts. In addition, we present a dataset that focuses on event-event dependency relationships.

The difficulty of planning languages lies in their need to precisely and efficiently describe complex planning issues, ensuring the coherence of formulated strategies. PDDL, a prominent planning language, is founded on the STRIPS approach and provides a structured method for describing planning problems via domain predicates and actions. An action in PDDL consists of three fundamental elements: parameters, preconditions, and effects, which map out the transition between states and are each represented by a set of predicates describing their properties.

The difficulties intensify when addressing actions in physical domains. Classical planning domains, such as the 8-Puzzle game, offer a simplified scenario in comparison to the physical domains' real-world complexities. Given the variables involved, an action such as shooting a projectile into a goal can have multiple outcomes, and state transitions result from significant changes. In an effort to address these complexities, we characterize states using qualitative spatial relations, a well-researched topic in AI. We present a novel model building technique that combines neighborhood and hierarchical clustering in an effort to group observed state transitions resulting from identical activities. Notably, our method also contributes to detecting and describing novelties in the physical domain, elucidating both their preconditions and effects, thereby facilitating the dissemination of knowledge among agents.

Narrative texts also presents a significant challenge. In contrast to structured data, narrative texts, which are rich in events, are by nature unstructured. In existing research, the extraction of action models from such texts has been primarily manual or semi-automated. We present NaRuto, a system designed to bridge this gap. In a two-step procedure, NaRuto derives structured events from narrative texts and then produces action models. We demonstrate the effectiveness of NaRuto by comparing it to examples from the literature and evaluating its action identification, action precondition/effects construction, and ability to generate alternative plans.

In narrative texts, events play a central role. While advancements have been made in relation extraction between entities, the interrelationships between events remain largely unexplored. We address this deficiency by introducing our human-annotated Event Dependency Relation (EDeR) dataset, which aims to extract event dependency information. Using the EDeR dataset, we refine the event representations of existing semantic role labeling models and apply them to a co-reference resolution task.

The automated generation and representation of planning action models, whether derived from physical domains or narrative texts, are crucial to AI planning. Through our methodologies, datasets, and systems, we intend to expedite these processes, resulting in more effective and precise AI planning applications across a variety of domains. Show more