A Silver Standard Corpus of Human Phenotype-Gene Relations

Proceedings of NAACL-HLT 2019, pages 1487–1492 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1487 A Silver Standard Corpus of Human Phenotype-Gene Relations Diana Sousa∗, Andre Lamurias and Francisco M. Couto LASIGE, Faculdade de Ciências, Universidade de Lisboa, Portugal Abstract Human phenotype-gene relations are funda- mental to fully understand the origin of some phenotypic abnormalities and their associated diseases. Biomedical literature is the most comprehensive source of these relations, how- ever, we need Relation Extraction tools to au- tomatically recognize them. Most of these tools require an annotated corpus and to the best of our knowledge, there is no corpus avail- able annotated with human phenotype-gene relations. This paper presents the Phenotype- Gene Relations (PGR) corpus, a silver stan- dard corpus of human phenotype and gene an- notations and their relations. The corpus con- sists of 1712 abstracts, 5676 human pheno- type annotations, 13835 gene annotations, and 4283 relations1. We generated this corpus us- ing Named-Entity Recognition tools, whose results were partially evaluated by eight cura- tors, obtaining a precision of 87.01%. By us- ing the corpus we were able to obtain promis- ing results with two state-of-the-art deep learn- ing tools, namely 78.05% of precision. The PGR corpus was made publicly available to the research community.2 1 Introduction Automatic extraction of relations between enti- ties mentioned in literature is essential to ob- tain knowledge that was already available but re- quired considerable manual effort and time to re- trieve. Recently, biomedical relation extraction has gained momentum in several text-mining ap- plications, such as event extraction and slot-filling (Lamurias and Couto, 2019b). Some of the com- monly extracted biomedical relations are protein- protein interactions (Papanikolaou et al., 2015), ∗dfsousa@lasige.di.fc.ul.pt 1Query 1, corresponds to the 10/12/2018 release of PGR 2https://github.com/lasigeBioTM/PGR drug-drug interactions (Lamurias et al., 2019) and disease-gene relationships (Kim et al., 2017). There are a few worth mention systems regard- ing biomedical Relation Extraction (RE) (Verga et al., 2018), and that specifically focus on the ex- traction of phenotype-gene relations regarding dif- ferent species types like plants (Xing et al., 2018) and humans (Collier et al., 2015). The main prob- lem that these systems face is a lack of specific high quality annotated corpora (gold standard cor- pus), mostly because this task requires not only a considerable amount of manual effort but also spe- cific expertise that is not widely available. A solu- tion to these limitations is to generate the corpus in a fully automated manner (silver standard corpus). Connecting human phenotypes to genes helps us to understand the origin of some phenotypic ab- normalities and their associated diseases. To ex- tract human phenotype-gene relations, both enti- ties, human phenotypes and genes have to be rec- ognized. With genes, as a result of lexical features being relatively regular, many systems can suc- cessfully identify them in text (Leaman and Gon- zalez, 2008). Even though Named-Entity Recog- nition (NER) research has significantly improved in the last years, human phenotype identification is still a complex task, …

Content truncated to 500 words