Explaining Relationships Between Scientific Documents

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2130–2144 August 1–6, 2021. ©2021 Association for Computational Linguistics 2130 Explaining Relationships Between Scientific Documents Kelvin Luu1,∗ Xinyi Wu1,∗ Rik Koncel-Kedziorski1 Kyle Lo2 Isabel Cachola3 Noah A. Smith1,2 1University of Washington 2Allen Institute for AI 3Johns Hopkins University {kellu,nasmith}@cs.washington.edu, {xywu,kedzior}@uw.edu, kylel@allenai.org, icachola@cs.jhu.edu Abstract We address the task of explaining relation- ships between two scientific documents using natural language text. This task requires mod- eling the complex content of long technical documents, deducing a relationship between these documents, and expressing that relation- ship in text. Successful solutions can help im- prove researcher efficiency in search and re- view. In this paper, we operationalize this task by using citing sentences as a proxy. We es- tablish a large dataset for our task. We pre- train a large language model to serve as the foundation for autoregressive approaches to the task. We explore the impact of taking dif- ferent views on the two documents, including the use of dense representations extracted with scientific information extraction systems. We provide extensive automatic and human evalu- ations which show the promise of such models, and make clear the challenges for future work. 1 Introduction The output of the world’s scientists doubles roughly every nine years (Bornmann and Mutz, 2015). Con- sequently, researchers must devote significant en- ergy to quickly understand how a new piece of research fits with a rapidly changing research land- scape. Several lines of research seek to reduce this bur- den on scientists. Citation recommendation sys- tems suggest references to relevant published work (McNee et al., 2002; Bhagavatula et al., 2018). In- tent classification systems help determine the type and importance of a citation in a work (Valenzuela et al., 2015; Cohan et al., 2019). Summarization systems aim to help researchers more quickly un- derstand the basic ideas in a piece of research (Co- han and Goharian, 2015; Yasunaga et al., 2019). We draw inspiration from these works as well as ∗Equal contribution. broader challenges like explaining the connection between concurrent works or relating a new paper to those a reader is already familiar with. Automatically describing inter-document rela- tionships could decrease the time researchers de- vote to literature review. For instance, explanations for a new paper can be personalized to a particu- lar reader by relating the new work to ones they have read before. Further, such technology could be incorporated into writing assistance systems to help less experienced or non-native writers better articulate the connection between their work and prior art. Additionally, users of citation recommen- dation systems can benefit from natural language explanations of recommendation system choices. In addition to the utility of this task to scientists, it presents several interesting technical challenges. These include effectively representing the impor- tant information in a document, generating from a long-tailed technical vocabulary, and expressing the variety of connections between related scien- tific papers. Figure 1 illustrates how the same …

Content truncated to 500 words