Extracting STRIPS Representations of Actions and Events

Proceedings of Recent Advances in Natural Language Processing, pages 1–8, Hissar, Bulgaria, 12-14 September 2011. Extracting STRIPS Representations of Actions and Events Avirup Sil and Alexander Yates Center for Data Analytics and Biomedical Informatics Temple University Broad St. and Montgomery Ave. Philadelphia, PA 19122 {avirup.sil, yates}@temple.edu Abstract Knowledge about how the world changes over time is a vital component of common- sense knowledge for Artificial Intelligence (AI) and natural language understanding. Actions and events are fundamental com- ponents to any knowledge about changes in the state of the world: the states before and after an event differ in regular and pre- dictable ways. We describe a novel sys- tem that tackles the problem of extracting knowledge from text about how actions and events change the world over time. We leverage standard language-processing tools, like semantic role labelers and coref- erence resolvers, as well as large-corpus statistics like pointwise mutual informa- tion, to identify STRIPS representations of actions and events, a type of representa- tion commonly used in AI planning sys- tems. In experiments on Web text, our ex- tractor’s Area under the Curve (AUC) im- proves by more than 31% over the clos- est system from the literature for identi- fying the preconditions and add effects of actions. In addition, we also extract sig- nificant aspects of STRIPS representations that are missing from previous work, in- cluding delete effects and arguments. 1 Introduction Common-sense knowledge about the changes in the state of the world over time is one of the most crucial forms of knowledge for an intelli- gent agent, since it informs an agent of the ways in which it can act upon the world. A recent survey of the common-sense knowledge involved in the recognizing textual entailment task demon- strates that knowledge about action and event se- mantics, in particular, constitutes a major compo- nent of the knowledge involved in understanding natural language (LoBue and Yates, 2011). This knowledge is also vital for central AI tasks like planning, plan recognition (Kautz, 1991; Geib and Steedman, 2007), and dialogue processing (Car- berry, 1990; Litman and Allen, 1987). In this paper we explore text mining approaches to extracting common-sense knowledge about ac- tion and event semantics. Our previous approach (Sil et al., 2010) (henceforth, S10) identifies the preconditions and effects of actions. We describe how we extend S10’s approach by identifying ad- ditional kinds of effects; by connecting this knowl- edge to an external ontology and generalizing the preconditions and effects; and by identifying ar- gument variables for each predicate. Our experi- ments show that our novel extractor can identify the fully-formed STRIPS representations of ac- tions with precision 0.73 and recall 0.72, and it improves on S10’s AUC for tasks that both sys- tems can handle by over 31%. The next section discusses previous work. Sec- tion 3 introduces STRIPS representation and the challenges involved in extracting such representa- tions. Section 4 details our extraction techniques. Section 5 presents our experiments. Section 6 con- cludes and discusses future work. …

Content truncated to 500 words