MedDistant19: Towards an Accurate Benchmark for Broad-Coverage Biomedical Relation Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 2259–2277 October 12–17, 2022. 2259 MEDDISTANT19: Towards an Accurate Benchmark for Broad-Coverage Biomedical Relation Extraction Saadullah Amin♣,△ ∗ Pasquale Minervini♠ ∗ David Chang♦ Pontus Stenetorp♠ Günter Neumann♣,△ ♣German Research Center for Artificial Intelligence ♠UCL Centre for Artificial Intelligence △Saarland Informatics Campus, Saarland University ♦Yale Center for Medical Informatics {saadullah.amin,guenter.neumann}@dfki.de {p.minervini,p.stenetorp}@cs.ucl.ac.uk david.chang@yale.edu Abstract Relation extraction in the biomedical domain is challenging due to the lack of labeled data and high annotation costs, needing do- main experts. Distant supervision is com- monly used to tackle the scarcity of anno- tated data by automatically pairing knowledge graph relationships with raw texts. Such a pipeline is prone to noise and has added chal- lenges to scale for covering a large num- ber of biomedical concepts. We investi- gated existing broad-coverage distantly super- vised biomedical relation extraction bench- marks and found a significant overlap between training and test relationships ranging from 26% to 86%. Furthermore, we noticed sev- eral inconsistencies in the data construction process of these benchmarks, and where there is no train-test leakage, the focus is on inter- actions between narrower entity types. This work presents a more accurate benchmark MEDDISTANT19 for broad-coverage distantly supervised biomedical relation extraction that addresses these shortcomings and is obtained by aligning the MEDLINE abstracts with the widely used SNOMED Clinical Terms knowl- edge base. Lacking thorough evaluation with domain-specific language models, we also con- duct experiments validating general domain re- lation extraction findings to biomedical rela- tion extraction. 1 Introduction Extracting structured knowledge from unstructured text is important for knowledge discovery and man- agement. Biomedical literature and clinical narra- tives offer rich interactions between entities men- tioned in the text (Craven and Kumlien, 1999; Xu and Wang, 2014), which can be helpful for applica- tions such as bio-molecular information extraction, pharmacogenomics, and identifying drug-drug in- teractions (DDIs), among others (Luo et al., 2017). ∗ Equal contribution. Iron deficiency is the most common MND worldwide and leads to microcytic anemia , decreased capacity for work , as well as impaired immune and endocrine function . Iron deficiency anaemia ( IDA ) and beta-thalassaemia are the most common causes of microcytic anaemia . Studies here reported indicated that the anemia is hypochromic and microcytic anemia of blood loss and iron deficiency , in spite of the presence of large amounts of iron in the pulmonary tissue . The high proportion of microcytic anaemia and the fact that gender differences were only seen after the menarche period in women suggest that iron deficiency was the main cause of anaemia . MCV/RBC and (MCV)2 X MCH separated successfully the subjects with microcytic anaemia ( heterozygous thalassaemia and iron deficiency ) from normal controls . Significantly higher serum homocysteine levels were reported in the iron deficiency anemia group compared to normal controls and in subjects with microcytic anemia and normal ferritin. CUI: (C0240066, C0085576) Semantic Type: (Disease or Syndrome, Disease or Syndrome) Semantic Group: (Disorders, Disorders) cause_of ✓ ✓ ✓ Figure 1: …

Content truncated to 500 words