An Attribution Relations Corpus for Political News

An Attribution Relations Corpus for Political News Edward Newell, Drew Margolin, Derek Ruths edward.newell@mail.mcgill.ca, dm658@corenell.edu, derek.ruths@mcgill.ca McGill University, Cornell University, McGill University Abstract An attribution occurs when an author quotes, paraphrases, or describes the statements and private states of a third party. Journalists use attribution to report statements and attitudes of public figures, organizations, and ordinary individuals. Properly recognizing attributions in context is an essential aspect of natural language understanding and implicated in many NLP tasks, but current resources are limited in size and completeness. We introduce the Political News Attribution Relations Corpus 2016 (PolNeAR)2—the largest, most complete attribution relations corpus to date. This dataset greatly increases the volume of high-quality attribution annotations, addresses shortcomings of existing resources, and expands the diversity of publishers sourced. PolNeAR is built on news articles covering the political candidates during the year leading up to US Presidential Election in November of 2016. The dataset will support the creation of sophisticated end-to-end solutions for attribution extraction and invite interdisciplinary collaboration between the NLP, communications, political science, and journalism communities. Along with the dataset we contribute revised guidelines aimed at improving clarity and consistency in the annotation task, and an annotation interface specially adapted to the task, for reproduction or extension of this work2. Keywords: attribution, quotation, sourcing, corpus, news, journalism, politics 1. Introduction Attribution occurs when an author describes a propositional attitude (Russel, 1940) held by some third party: an agent’s statements, intentions, beliefs, knowledge, perceptions, de- crees, or sentiments about something (see Table 1). Quot- ing or paraphrasing another person is a familiar form of attribution. But self-attribution, and attribution to artifacts like reports, recordings, or databases are included in the definition. Prior work defines attributions as consisting of three parts: (1) the source, to whom content is attributed; (2) the con- tent that is attributed; and (3) a cue phrase used to signal attribution, such as “said” or “according to” (Pareti, 2012). While the statements made by public figures are often in- herently newsworthy, attribution to any source is a funda- mental mechanism for journalists to lend credibility or au- thority to, or to nuance, an assertion. Accurate and infor- mative attributions provide readers with transparency and accountability by making journalists’ sourcing identifiable (Esser and Umbricht, 2014). Lately, there has been increased skepticism directed at the mainstream media1, with questions about the legitimacy of reporting often focusing on sourcing. Thus, attribution phe- nomena are of fundamental interest for the maintenance of journalistic standards, and should be carefully attended to by the critical reader. Although attribution is a fundamental rhetorical mechanism in media, it has not received much attention from compu- tational researchers, and partly as a result, datasets for its study are limited in number and depth. Recently, however, direct efforts have been made to study attribution, with the creation of PARC3 (Pareti, 2012), the largest attribution- relations dataset prior to PolNeAR. Despite its important contribution, PARC3 suffers from low annotator recall— many attributions have gone unnoticed by annotators. This impedes the creation …

Content truncated to 500 words