Relational Summarization for Corpus Analysis

Proceedings of NAACL-HLT 2018, pages 1760–1769 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Relational Summarization for Corpus Analysis Abram Handler and Brendan O’Connor College of Information and Computer Sciences University of Massachusetts Amherst ahandler@cs.umass.edu, brenocon@cs.umass.edu Abstract This work introduces a new problem, rela- tional summarization, in which the goal is to generate a natural language summary of the re- lationship between two lexical items in a cor- pus, without reference to a knowledge base. Motivated by the needs of novel user inter- faces, we define the task and give examples of its application. We also present a new query- focused method for finding natural language sentences which express relationships. Our method allows for summarization of more than two times more query pairs than baseline re- lation extractors, while returning measurably more readable output. Finally, to help guide future work, we analyze the challenges of re- lational summarization using both a news and a social media corpus. 1 Introduction Research on automatic summarization (Nenkova et al., 2011; Das and Martins, 2007) aims to help users understand large document sets. However, the details of how textual summaries might actu- ally be presented to users are often ignored. We propose that user interfaces which display note- worthy terms or concepts present the need for re- lational summaries: descriptions of the relation- ship between two entities or noun phrases from a corpus. Examples of such interfaces include: comman- dline software for examining noteworthy terms or phrases (Squirrell, 2017; Robinson, 2016; Mon- roe et al., 2008), point-and-click browsers which display named entities and their interconnec- tions on a network diagram (Wright et al., 2009; G¨org et al., 2014; Tannier, 2016), concept map browsers (Falke and Gurevych, 2017b) and doc- ument search engines which suggest terms rele- vant to a query, such as the related searches dis- played on Wikipedia info boxes from Google. In Aristide Aristide Gen. Cedras UN Liberation Theology rival of influenced by relied on Aristide the Haitian leader … governing philosophy informed by liberation theology Aristide, as a young Catholic priest … was influenced by the liberation theology Aristide was earlier expelled from Salesian Order for promoting liberation theology Clinton criticized concept map snippet box Figure 1: An example interface which requires rela- tional summarization. The user has queried for the en- tity Aristide. The interface shows a concept map (top), displaying short summaries of Aristide’s important re- lationships. The user has drilled down to see a more detailed summary of Aristide’s relationship with liber- ation theology, displayed in a snippet box (bottom). all such settings a natural question arises: what is the nature of the relationship between the entities or concepts shown in the interface? One particular interface which presents the need for a relational summary is shown in figure 1. Relational questions are ubiquitous and varied. Examples include the following. What is the rela- tionship between the “City of London” and “goal- delivery of Newgate” in 18th century court records (Hitchcock et al., 2012)? …

Content truncated to 500 words