Towards Learning Operator Schema from Free Text

In automated planning current research is focused on developing domain-independent planning engines. These require domain models, written in a standard input language such as PDDL to supply knowledge of the planning application and task, before they can be used. The main component of a domain model is the representation of actions in the form of lifted opera- tor schema. The acquisition and engineering of these schema is an important area of research, as this process is recognised as being di�cult and laborious even for planning experts. A fruitful line of research is to investigate mechanisms to automatically learn planning domain models. Re- cent research has studied learning from structured or re�ned inputs supplied by a training agent (Cress- well, McCluskey, and West 2011; Zhuo et al. 2010; Wu, Yang, and Jiang 2005; McCluskey et al. 2010). An alternative method would be to allow planning agents to learn and develop the domain models by observa- tion. One freely available source for learning actions is selected web text; here actions are represented as verbs in natural language. This project aims to in- vestigate the possibility of extracting formal structures representing actions from free text. We intend to utilise large text corpuses available on-line from which to ex- tract such action knowledge, and learn operator schema in a formal language that can be converted to PDDL.