An Analytical Study on a Benchmark Corpus Constructed for Related Work Generation

Automatic related work generation aims at producing a related work section for a given scientific paper. Demand for this task replacing a labor-intensive process has substantially increased in recent years. Considering the lack of an open and large-scale dataset for related work generation, we introduce NudtRwG (https://​github.​com/​NudtRwG/​NudtRwG-Dataset/​), a collection of 2,084 document sets, each with a target paper, a ground truth related work, and the corresponding reference papers. To our knowledge, NudtRwG is the first open, large-scale and high-quality dataset for related work generation. The contribution of this work apart from the dataset is two-fold: firstly, we present a detailed description of the data collection procedure along with an analysis on the characteristics of the dataset; secondly, we conduct an analytical study, investigating the effects of summative sections (abstract, introduction and conclusion) and other sections of reference papers on related work generation. Experiments reveal that the two parts are equally important and other sections should not be ignored. When generating a related work section, researchers should consider not only summative sections, but also other sections of reference papers.