MultiCite: Modeling realistic citations requires moving beyond the single-sentence single-label setting

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1875 - 1889 July 10-15, 2022 ©2022 Association for Computational Linguistics MULTICITE: Modeling realistic citations requires moving beyond the single-sentence single-label setting Anne Lauscher∗1 Brandon Ko2 Bailey Kuehl3 Sophie Johnson3 Arman Cohan3 David Jurgens4 Kyle Lo3 1MilaNLP, Bocconi University, Italy, 2University of Washington, Seattle WA, 3AI2, Seattle WA, 4School of Information, University of Michigan, Ann Arbor MI 1 anne.lauscher@unibocconi.it, 2bk36@cs.washington.edu, 3{armanc,baileyk,sophiej,kylel}@allenai.org, 4jurgens@umich.edu Abstract Citation context analysis (CCA) is an impor- tant task in natural language processing that studies how and why scholars discuss each oth- ers’ work. Despite decades of study, com- putational methods for CCA have largely re- lied on overly-simplistic assumptions of how authors cite, which ignore several important phenomena. For instance, scholarly papers often contain rich discussions of cited work that span multiple sentences and express mul- tiple intents concurrently. Yet, recent work in CCA is often approached as a single-sentence, single-label classification task, and thus many datasets used to develop modern computational approaches fail to capture this interesting dis- course. To address this research gap, we high- light three understudied phenomena for CCA and release MULTICITE, a new dataset of 12.6K citation contexts from 1.2K computa- tional linguistics papers that fully models these phenomena. Not only is it the largest col- lection of expert-annotated citation contexts to-date, MULTICITE contains multi-sentence, multi-label citation contexts annotated through- out entire full paper texts. We demonstrate how MULTICITE can enable the development of new computational methods on three important CCA tasks. We release our code and dataset at https://github.com/allenai/multicite. 1 Introduction Citations connect the current paper to the broader discourse of science (e.g., Garfield, 1955; Sid- dharthan and Teufel, 2007), help signal future im- pact and uses (e.g., McKeown et al., 2016), and, in downstream applications, can aid in summarizing a work’s contributions (e.g., Qazvinian and Radev, 2008; Cohan and Goharian, 2015; Lauscher et al., 2017a). The study of the role and purpose of cita- tions, known as citation context analysis (CCA ∗ Part of the work was conducted during an internship at the Allen Institute for AI and while being affiliated to the University of Mannheim. Swales, 1986), has uncovered traces of how ideas have influenced a field, collaboration and competi- tion among peers, and trends in scientific fields. Computational approaches to CCA have largely focused on classifying the intent or purpose of a particular citation. With the advent of deep neu- ral models, recent CCA research efforts have fo- cused on increasing the size of the published re- sources (Cohan et al., 2019; Tuarob et al., 2019; Pride and Knoth, 2020). However, larger data sets have come at the expense of oversimplifying the rich discourse patterns surrounding citations, espe- cially as complexity of annotation is often traded for large-scale data collection. The aforementioned recent large-scale resources, for instance, have ap- proached CCA as a single-sentence, single-label classification task, overlooking the richness and nuance with which citations are used …

Content truncated to 500 words