CORWA: A Citation-Oriented Related Work Annotation Dataset

TL;DR

CORWA addresses the problem of generating related work sections by introducing a linguistically motivated annotation scheme that decomposes related work into discourse roles, citation spans, and citation types. The dataset enables multi-task modeling via a joint tagger that uses a SciBERT-based encoder and task-specific decoders, achieving strong token-level F1 across tasks and enabling distantly supervised expansion to unlabeled data. The authors demonstrate that citation spans, rather than full sentences, are a more effective target for retrieval and span-based generation, and they present a span-based generation pilot using Longformer variants with encouraging human judgments. They further advocate a human-in-the-loop, iterative framework for full related work generation to mitigate long-context generation challenges and allow user-guided refinement. Overall, CORWA provides a scalable, reusable resource and modeling blueprint for span- and discourse-aware related work generation with practical implications for automatic literature reviews and scholarly writing support.

Abstract

Academic research is an exploratory activity to discover new solutions to problems. By this nature, academic research works perform literature 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 generation framework, we present a Citation Oriented Related Work Annotation (CORWA) dataset that labels different types of citation text fragments 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, iterative, abstractive related work generation.

Figures (6)

• Figure 1: An example of CORWA labels displayed using the BRAT interface stenetorp2012brat.
• Figure 2: Histogram of the length of dominant and reference-type citation spans, excluding citation marks. The dashed vertical lines are the means of dominant and reference span lengths, 34.5 and 8.2, respectively.
• Figure 3: Parallel plot of the proportion of summarization and narrative sentences in each paragraph. Paragraphs with neither type of sentences are excluded.
• Figure 4: The architecture of our joint related work tagger, which performs discourse tagging (Disc), citation type recognition (CT), and citation span detection (CS).
• Figure 5: Histogram of top-1 ROUGE recall scores of retrieved sentences from cited papers using different queries. The dashed vertical lines are the means of reference sentence (0.220), dominant sentence (0.293), dominant span (0.316), and reference spans (0.449).