A Deep Metric Learning Method for Biomedical Passage Retrieval

Proceedings of the 28th International Conference on Computational Linguistics, pages 627–637 Barcelona, Spain (Online), December 8-13, 2020 627 Named Entity Recognition for Chinese biomedical patents Yuting Hu LIACS Leiden University, the Netherlands yukihuyt@gmail.com Suzan Verberne LIACS Leiden University, the Netherlands s.verberne@liacs.leidenuniv.nl Abstract There is a large body of work on Biomedical Entity Recognition (Bio-NER) for English, but there have only been a few attempts addressing NER for Chinese biomedical texts. Because of the growing amount of Chinese biomedical discoveries being patented, and lack of NER models for patent data, we train and evaluate NER models for the analysis of Chinese biomedical patent data, based on BERT. By doing so, we show the value and potential of this domain-specific NER task. For the evaluation of our methods we built our own Chinese biomedical patents NER dataset, and our optimized model achieved an F1 score of 0.54±0.15. Further biomedical analysis indicates that our solution can help detecting meaningful biomedical entities and novel gene–gene interactions, with limited labeled data, training time and computing power. 1 Introduction In the area of biomedical NLP, Named Entity Recognition (NER) is a widely discussed and studied topic. The aim of the task is to identify biomedical entities such as genes, proteins, cell types, and diseases in biomedical documents, to allow for knowledge discovery in this domain. Models for Biomedical NER (Bio-NER) offer the opportunity to mine information and knowledge and thereby foster biomedical and drug discovery research (Habibi et al., 2017). Several shared tasks addressing Bio-NER have been organized. These attempts and tasks resulted in benchmark datasets for solving English Bio-NER tasks, e.g. the GENIA corpus (Kim et al., 2003), JNLPBA (Kim et al., 2004), and BC2GM (Smith et al., 2008). However, when turning our attention to Chinese bio-NER, only a few attempts have been made, and these attempts either had limitations in text resource types and amount (Gu et al., 2008) or did not pre- define biomedical named entity categories (Wang et al., 2019). Another limitation is that these attempts mostly focus on clinical texts or biomedical scientific publications but not include other relevant text resources, in particular biomedical patents. Many biomedical discoveries are patented in China, not only because of the encouraging policy on patentability of genetic products, but also since the existence of the speedily progressed and cheaper gene sequencing services (Du, 2018). Patent texts are highly technical with long sentences (Verberne et al., 2010). Two additional challenges of Chinese biomedical patents that we encountered are OCR errors and the heavy usage of code-mixing expressions, mixing English and Chinese in one entity. This is mainly because the protein and gene names are commonly written in English (or the English names are given after the Chinese ones), while the disease names and other contents are written in Chinese. For this reason, it is not possible to directly apply pre-trained NLP models to Chinese biomedical patents. Moreover, as mentioned before, because of the lack of related studies, we not only lack pre-trained models which were trained …

Content truncated to 500 words