Two key obstacles in biomedical relation extraction (RE) are the scarcity of annotations and the prevalence of instances without explicitly pre-defined labels due to low annotation coverage. Existing approaches, which treat biomedical RE as a multi-class classification task, often result in poor generalization in low-resource settings and do not have the ability to make selective prediction on unknown cases but give a guess from seen relations, hindering the applicability of those approaches. We present NBR, which converts biomedical RE as natural language inference formulation through indirect supervision. By converting relations to natural language hypotheses, NBR is capable of exploiting semantic cues to alleviate annotation scarcity. By incorporating a ranking-based loss that implicitly calibrates abstinent instances, NBR learns a clearer decision boundary and is instructed to abstain on uncertain instances. Extensive experiments on three widely-used biomedical RE benchmarks, namely ChemProt, DDI and GAD, verify the effectiveness of NBR in both full-set and low-resource regimes. Our analysis demonstrates that indirect supervision benefits biomedical RE even when a domain gap exists, and combining NLI knowledge with biomedical knowledge leads to the best performance gains.
Lack of supervision results in poor low-resource #InformationExtraction performance, especially for #biomedical domain with limited annotations. In our #ACL2023 main paper, we introduce NBR to enhance #biomedical #RelationExtraction with #IndirectSupervision from the NLI task.1/5 pic.twitter.com/f2MtDf7bmr— Mingyu Derek MA (@mingyu_ma) July 8, 2023
NBR reformulates #RelationExtraction to #NaturalLanguageInference by treating the input as the premise while verbalizing each relation label into hypotheses, to utilize the supervision from the high-resource general-domain NLI task for the resource-hungry biomedical RE task. 2/5— Mingyu Derek MA (@mingyu_ma) July 8, 2023
We propose a ranking-based loss that implicitly handles abstinent relations ubiquitous in biomedical RE by contrastively calibrating the score of abstinent instances. We introduce an explicit abstention detector that specializes in "has relation" vs "no relation" detection. 3/5— Mingyu Derek MA (@mingyu_ma) July 8, 2023
Experiments on ChemProt, DDI and GAD show that NBR provides consistent improvements across the board in high-resource settings (left figure) and up to 34 points F1 score improvement for low-resource #biomedical #RelationExtraction (right figure). 4/5 pic.twitter.com/scgdIP2DgI— Mingyu Derek MA (@mingyu_ma) July 8, 2023
This work is done with the fantastic @JiashuXu2 @muhao_chen— Mingyu Derek MA (@mingyu_ma) July 8, 2023
There will be an oral presentation on Tuesday July 11th Session 5 17:15-17:30 at the Metropolitan Centre room #ACL2023 @aclmeeting. See you soon! 5/5 pic.twitter.com/WYA7PuPEBr