Unsupervised Learning of Narrative Event Chains

Proceedings of ACL-08: HLT, pages 789–797, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Unsupervised Learning of Narrative Event Chains Nathanael Chambers and Dan Jurafsky Department of Computer Science Stanford University Stanford, CA 94305 {natec,jurafsky}@stanford.edu Abstract Hand-coded scripts were used in the 1970-80s as knowledge backbones that enabled infer- ence and other NLP tasks requiring deep se- mantic knowledge. We propose unsupervised induction of similar schemata called narrative event chains from raw newswire text. A narrative event chain is a partially ordered set of events related by a common protago- nist. We describe a three step process to learn- ing narrative event chains. The first uses unsu- pervised distributional methods to learn narra- tive relations between events sharing corefer- ring arguments. The second applies a tempo- ral classifier to partially order the connected events. Finally, the third prunes and clusters self-contained chains from the space of events. We introduce two evaluations: the narrative cloze to evaluate event relatedness, and an or- der coherence task to evaluate narrative order. We show a 36% improvement over baseline for narrative prediction and 25% for temporal coherence. 1 Introduction This paper induces a new representation of struc- tured knowledge called narrative event chains (or narrative chains). Narrative chains are partially or- dered sets of events centered around a common pro- tagonist. They are related to structured sequences of participants and events that have been called scripts (Schank and Abelson, 1977) or Fillmorean frames. These participants and events can be filled in and instantiated in a particular text situation to draw in- ferences. Chains focus on a single actor to facili- tate learning, and thus this paper addresses the three tasks of chain induction: narrative event induction, temporal ordering of events and structured selection (pruning the event space into discrete sets). Learning these prototypical schematic sequences of events is important for rich understanding of text. Scripts were central to natural language understand- ing research in the 1970s and 1980s for proposed tasks such as summarization, coreference resolu- tion and question answering. For example, Schank and Abelson (1977) proposed that understanding text about restaurants required knowledge about the Restaurant Script, including the participants (Cus- tomer, Waiter, Cook, Tables, etc.), the events consti- tuting the script (entering, sitting down, asking for menus, etc.), and the various preconditions, order- ing, and results of each of the constituent actions. Consider these two distinct narrative chains. accused X W joined X claimed W served X argued W oversaw dismissed X W resigned It would be useful for question answering or tex- tual entailment to know that ‘X denied ’ is also a likely event in the left chain, while ‘ replaces W’ temporally follows the right. Narrative chains (such as Firing of Employee or Executive Resigns) offer the structure and power to directly infer these new subevents by providing critical background knowl- edge. In part due to its complexity, automatic in- duction has not been addressed since the early non- statistical work of Mooney and DeJong (1985). The …

Content truncated to 500 words