Unsupervised Learning of Narrative Schemas and their Participants

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 602–610, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Unsupervised Learning of Narrative Schemas and their Participants Nathanael Chambers and Dan Jurafsky Stanford University, Stanford, CA 94305 {natec,jurafsky}@stanford.edu Abstract We describe an unsupervised system for learn- ing narrative schemas, coherent sequences or sets of events (arrested(POLICE,SUSPECT), convicted( JUDGE, SUSPECT)) whose arguments are filled with participant semantic roles defined over words (JUDGE = {judge, jury, court}, POLICE = {police, agent, authorities}). Unlike most previous work in event structure or semantic role learning, our sys- tem does not use supervised techniques, hand-built knowledge, or predefined classes of events or roles. Our unsupervised learning algorithm uses corefer- ring arguments in chains of verbs to learn both rich narrative event structure and argument roles. By jointly addressing both tasks, we improve on pre- vious results in narrative/frame learning and induce rich frame-specific semantic roles. 1 Introduction This paper describes a new approach to event se- mantics that jointly learns event relations and their participants from unlabeled corpora. The early years of natural language processing (NLP) took a “top-down” approach to language understanding, using representations like scripts (Schank and Abelson, 1977) (structured represen- tations of events, their causal relationships, and their participants) and frames to drive interpreta- tion of syntax and word use. Knowledge structures such as these provided the interpreter rich infor- mation about many aspects of meaning. The problem with these rich knowledge struc- tures is that the need for hand construction, speci- ficity, and domain dependence prevents robust and flexible language understanding. Instead, mod- ern work on understanding has focused on shal- lower representations like semantic roles, which express at least one aspect of the semantics of events and have proved amenable to supervised learning from corpora like PropBank (Palmer et al., 2005) and Framenet (Baker et al., 1998). Un- fortunately, creating these supervised corpora is an expensive and difficult multi-year effort, requiring complex decisions about the exact set of roles to be learned. Even unsupervised attempts to learn semantic roles have required a pre-defined set of roles (Grenager and Manning, 2006) and often a hand-labeled seed corpus (Swier and Stevenson, 2004; He and Gildea, 2006). In this paper, we describe our attempts to learn script-like information about the world, including both event structures and the roles of their partic- ipants, but without pre-defined frames, roles, or tagged corpora. Consider the following Narrative Schema, to be defined more formally later. The events on the left follow a set of participants through a series of con- nected events that constitute a narrative: A search B A arrest B D convict B B plead C D acquit B D sentence B A = Police B = Suspect C = Plea D = Jury Events Roles Being able to robustly learn sets of related events (left) and frame-specific role information about the argument types that fill them (right) could assist a variety of NLP applications, from question …

Content truncated to 500 words