Improving Distantly Supervised Relation Extraction by Natural Language Inference

Authors: Kang Zhou, Qiao Qiao, Yuepei Li, Qi Li

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that the proposed framework significantly improves the SOTA performance (up to 7.73% of F1) on distantly supervised RE benchmark datasets.
Researcher Affiliation Academia Kang Zhou, Qiao Qiao, Yuepei Li, Qi Li Department of Computer Science, Iowa State University, Ames, Iowa, USA {kangzhou, qqiao1, liyp0095, qli}@iastate.edu
Pseudocode Yes Algorithm 1: SARV
Open Source Code Yes Our code is available at https://github.com/kangISU/DSRE-NLI.
Open Datasets Yes We conduct experiments to evaluate the proposed framework on the widely-used public dataset: New York Times (NYT), which is a large-scale distantly labeled dataset constructed from NYT corpus using Freebase as the distant supervision (Riedel, Yao, and Mc Callum 2010). Recently, Jia et al. (2019) manually labeled a subset of the data as the testing set for a more accurate evaluation and constructed two versions of the dataset: NYT10.1 and NYT10.21. The statistics of the two datasets are summarized in Table 1, and more details about the instance generation can be found in Technical Appendix Section 1.
Dataset Splits Yes Table 1: Statistics of used datasets. Dev # total inst 2,379 (NYT10.1), 4,569 (NYT10.2), 22,631 (TACREV)
Hardware Specification Yes We run all methods using one Tesla V100S GPU (32G).
Software Dependencies No The paper mentions using "the pretrained De BERTa v2 model (He et al. 2020a)" but does not specify version numbers for general software dependencies like Python, PyTorch, or specific libraries used for implementation.
Experiment Setup Yes We train DSRE-NLI for 2 epochs on both NYT10.1 and NYT10.2 training variants. ... set the entailment probability threshold τ to 0.95