Learning task hierarchies using statistical semantics and goal reasoning

This paper describes WORD2HTN, an algorithm for learning hierarchical tasks and goals from plan traces in planning domains. WORD2HTN combines semantic text analysis techniques and subgoal learning in order to generate Hierarchical Task Networks (HTNs). Unlike existing HTN learning algorithms, WORD2HTN learns distributed vector representations that represent the similarities and semantics of the components of plan traces. WORD2HTN uses those representations to cluster them into task and goal hierarchies, which can then be used for automated planning. We describe our algorithm and present our preliminary evaluation thereby demonstrating the promise of WORD2HTN.