Numeric Information Extraction from Dataset Descriptions (2025)

NIED: A Corpus for Numeric Information Extraction from Dataset Descriptions

Abstract: Although a large number of machine learning datasets have been proposed, their descriptions are often insufficiently structured, making it difficult to search datasets using quantitative criteria. To address this issue, we construct NIED, an annotated corpus consisting of 3,926 dataset descriptions collected from academic papers and data repositories, designed for the task of extracting numerical information from dataset descriptions. In addition, we propose a two-stage labeling scheme that distinguishes numerical entities from contextual non-numerical information.

Authors: Moriyuki Kamoto, Akihiro Tamura, Marie Katsurai

Publication venue: JCDL 2025

Corpus

• GitHub