Modeling Document-level Causal Structures for Event Causal Relation Identification

Proceedings of NAACL-HLT 2019, pages 1808–1817 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1808 Modeling Document-level Causal Structures for Event Causal Relation Identification Lei Gao, Prafulla Kumar Choubey, Ruihong Huang Department of Computer Science and Engineering Texas A&M University (sjtuprog, prafulla.choubey, huangrh)@tamu.edu Abstract We aim to comprehensively identify all the event causal relations in a document, both within a sentence and across sentences, which is important for reconstructing pivotal event structures. The challenges we identified are two: 1) event causal relations are sparse among all possible event pairs in a docu- ment, in addition, 2) few causal relations are explicitly stated. Both challenges are espe- cially true for identifying causal relations be- tween events across sentences. To address these challenges, we model rich aspects of document-level causal structures for achiev- ing comprehensive causal relation identifica- tion. The causal structures include heavy in- volvements of document-level main events in causal relations as well as several types of fine- grained constraints that capture implications from certain sentential syntactic relations and discourse relations as well as interactions be- tween event causal relations and event corefer- ence relations. Our experimental results show that modeling the global and fine-grained as- pects of causal structures using Integer Lin- ear Programming (ILP) greatly improves the performance of causal relation identification, especially in identifying cross-sentence causal relations. 1 Introduction Understanding causal relations between events in a document is an important step in text under- standing and is beneficial to various NLP appli- cations, such as information extraction, question answering and text summarization. Causal rela- tions can occur between any two events in a doc- ument, both between events within a sentence and between events across sentences. In this paper, we aim to identify all the event causal relations in a document. The main challenges for achieving comprehen- sive causal relation identification are that event Figure 1: An Example of Main Event Causal Structure causal relations are sparse among all the event pairs in a document and few event causal relations are explicitly stated. The challenges are especially true for identifying cross-sentence event causal re- lations and most of them have no clear causal indi- cators. To address these challenges, we model rich aspects of document-level causal structures, i.e., structural distributions of causal relations within a document, for achieving comprehensive causal re- lation identification in news articles. Our key observation for improving causal re- lation identification is that causal relations, espe- cially cross-sentence causal relations, tend to in- volve one or two main events of a document. The main events are the focus of a story, which are usu- ally mentioned in the title of an article and have repeated mentions throughout the document. In- tuitively, causal relations in a document are often used to explain why the main events happened as well as consequences of the main events. For ex- ample, as shown in figure 1, killing is the main event. The events crossfire, spraying, richo- cheted, struck are its preconditions, and …

Content truncated to 500 words