Improved Induction of Narrative Chains via Cross-Document Relations

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, pages 208 - 212 July 14-15, 2022 ©2022 Association for Computational Linguistics Improved Induction of Narrative Chains via Cross-Document Relations Andrew Blair-Stanek University of Maryland Carey School of Law ablair-stanek@law.umaryland.edu Benjamin Van Durme Department of Computer Science Johns Hopkins University vandurme@cs.jhu.edu Abstract The standard approach for inducing narrative chains considers statistics gathered per individ- ual document. We consider whether statistics gathered using cross-document relations can lead to improved chain induction. Our study is motivated by legal narratives, where cases typ- ically cite thematically similar cases. We con- sider four novel variations on pointwise mutual information (PMI), each accounting for cross- document relations in a different way. One proposed PMI variation performs 58% better relative to standard PMI on recall@50 and in- duces qualitatively better narrative chains. 1 Introduction Narrative chains are sets of events centered around a common protagonist. They can be induced from corpora using various unsupervised methods, many using pointwise mutual information (PMI) between events. To our knowledge, no prior work has used the information available in relations between doc- uments in a corpus when inducing narrative chains. To illustrate the potential for improved narrative chain induction based on document relations, we develop four novel variants of pointwise mutual information (PMI) that assume a directed graph structure between documents (i.e. relations that are edges). We test these1 on a corpus of all U.S. federal court cases, which has a readily accessible relation between documents: citation. One case will cite prior cases as precedent in explaining its decision. We find that one of our four variants of PMI performs particularly well in the standard event cloze evaluation (Chambers and Jurafsky, 2008) and in inducing meaningful narrative chains. 2 Background Unsupervised narrative chain induction from a corpus was introduced by Chambers and Jurafsky 1Code is at https://github.com/BlairStanek/ cross-doc (2008), inspired by the notion of scripts owing to Schank and Abelson (1977). Coreference chains were extracted over the Gigaword corpus (Graff, 2002) to extract event chains with the same pro- tagonist. A syntactic parser identified each event in which the protagonist was involved, defined as the combination of a verb and dependency relation, such as (convict, obj). They then calculated the pointwise mutual information (PMI) (Church and Hanks, 1989, 1990) for each combination of two events and used this PMI to do agglomerative clus- tering to induce narrative chains. We follow the basic approach of Chambers and Jurafsky (2008), with the major extension that we take relations be- tween the documents into account for the first time. There have been numerous improvements on the Chambers and Jurafsky (2008) approach, including using language modeling approaches (Rudinger et al., 2015), neural networks (Pichotta and Mooney, 2016; Weber et al., 2018), and graphs where events are the vertices (Li et al., 2018, 2021). None of these improvements has considered rela- tions between documents in the corpus. 3 Alternative Measures Much of the narrative chain induction literature, following Chambers and Jurafsky (2008), has used …

Content truncated to 500 words