Unsupervised Learning of Narrative Event Chains

This paper addresses the problem of unsupervised extraction of narrative event chains, which are partially ordered sets of narrative events centered around a common protagonist. To extract narrative event chains, the paper assumes two things:

  • narrative coherence: verbs sharing co-referring arguments are semantically connected
  • the protagonist: although a narrative has several participants, there is a central actor who characterizes a narrative chain, which is the protagonist

Narrative chains are somewhat related to structured sequences of participants and events that are called scripts (Schank and Abelson, 1977).

As covered in the guest lecture by Brendan O'Connor, scripts were central to natural language understanding research in the 1970s and 1980s for tasks such as question answering, story understanding, summarization, etc. For example, Schank and Abelson (1977)'s restaurant script which is proposed to understand text about restaurants based on the participants (customer, waiter, etc.), their actions or movements (entering, sitting down, ordering, ingesting, etc), and effects resulting from each actions which can be seen as a collection of events with some sort of (temporal) ordering. These scripts or template (or "sketchy scripts" in the case of DeJong's FRUMP system, 1982) were manually written. The purpose of this paper is to learn such "scripts" from a collection of documents automatically.

The objective of this paper is to automatically extract narrative event chains, which are collections of events that are sharing a common protagonist and are temporally ordered.

The paper proposes three steps process to learning narrative event chains:

OpenSherlock Project »OpenSherlock Project
References »References
Conference Papers »Conference Papers
Unsupervised Learning of Narrative Event Chains
+Comments (0)
+Citations (1)