CORWA: A Citation-Oriented Related Work Annotation Dataset

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5426 - 5440 July 10-15, 2022 ©2022 Association for Computational Linguistics CORWA: A Citation-Oriented Related Work Annotation Dataset Xiangci Li Biswadip Mandal Jessica Ouyang Department of Computer Science University of Texas at Dallas Richardson, TX 75080 lixiangci8@gmail.com, {Biswadip.Mandal, Jessica.Ouyang}@UTDallas.edu Abstract Academic research is an exploratory activity to discover new solutions to problems. By this nature, academic research works perform litera- ture reviews to distinguish their novelties from prior work. In natural language processing, this literature review is usually conducted under the “Related Work” section. The task of related work generation aims to automatically generate the related work section given the rest of the research paper and a list of papers to cite. Prior work on this task has focused on the sentence as the basic unit of generation, neglecting the fact that related work sections consist of variable length text fragments derived from different information sources. As a first step toward a linguistically-motivated related work genera- tion framework, we present a Citation Oriented Related Work Annotation (CORWA) dataset that labels different types of citation text frag- ments from different information sources. We train a strong baseline model that automatically tags the CORWA labels on massive unlabeled related work section texts. We further suggest a novel framework for human-in-the-loop, iter- ative, abstractive related work generation. 1 Introduction Academic research is an exploratory activity to solve problems that have never been solved before. By this nature, each academic research work must sit at the frontier of its field and present novel contri- butions that have not been addressed by prior work; in order to convince readers of the novelty of the current work, the authors must compare against the prior work. While the format may vary among dif- ferent fields, in natural language processing (NLP), this literature review is usually conducted under the “Related Work” section. Since each paper must review the relevant prior work in its field, which is shared among papers on the same topic or task, many related work sections in a given field can be similar in both content and format. Therefore, Figure 1: An example of CORWA labels displayed using the BRAT interface (Stenetorp et al., 2012). it is a natural motivation to develop a system for generating related work sections automatically. The task of automatic related work generation is that of generating the related work section of a target paper given the rest of the target paper and a set of papers to cite. Prior works (Hoang and Kan, 2010; Hu and Wan, 2014; Chen and Zhuge, 2019; Wang et al., 2019; Xing et al., 2020; Ge et al., 2021; Luu et al., 2021; Chen et al., 2021) mostly simplify related work generation as a general summariza- tion task, generating related work sections using sentence-level models. This approach ignores the nature of the related work section, which consists of variable-length text fragments derived from differ- ent …

Content truncated to 500 words