Combinatory Categorial Grammar Learning for Plan Recognition in Domains with Type Trees
Combinatory Categorial Grammars (CCGs) have been used for plan recognition and planning in various domains such as robotics and video games. Prior work on CCG planning and plan recognition used hand-authored CCGs, which requires domain knowledge, and can be both time-consuming and error prone to construct. This paper focuses on automatically learning CCGs and presents two key contributions to the literature of CCG learning. First, we extend existing CCG learning algorithms for domains with predefined type trees via searching over type trees to find type generalizations. Second, we present the first comparison of learned CCGs with CCGs hand-authored by humans, showing tradeoffs of these algorithms. We apply this extended learning algorithm to Minecraft and Monroe, and demonstrate its performance against a hand-authored model for plan recognition tasks.