Learning Entity-Likeness with Multiple Approximate Matches for Biomedical NER

Proceedings of Recent Advances in Natural Language Processing, pages 1040–1049 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_117 1040 Learning Entity-Likeness with Multiple Approximate Matches for Biomedical NER An Nguyen Le, Hajime Morita and Tomoya Iwakura Artificial Intelligence Laboratory, Fujitsu Research, Fujitsu Ltd. 4 Chome-1-1 Kamikodanaka, Nakahara Ward, Kawasaki, Kanagawa, Japan {nguyenle.an,hmorita,iwakura.tomoya}@fujitsu.com Abstract Biomedical Named Entities are complex, so approximate matching has been used to im- prove entity coverage. However, the usual approximate matching approach fetches only one matching result, which is often noisy. In this work, we propose a method for biomed- ical NER that fetches multiple approximate matches for a given phrase to leverage their variations to estimate entity-likeness. The model uses pooling to discard the unnecessary information from the noisy matching results, and learn the entity-likeness of the phrase with multiple approximate matches. Experimental results on three benchmark datasets from the biomedical domain, BC2GM, NCBI-disease, and BC4CHEMD, demonstrate the effective- ness. Our model improves the average F- measures by up to 0.21 percentage points com- pared to a BioBERT-based NER. 1 Introduction In the biomedical field, obtaining labelled data is very costly. Biomedical Named Entities (NEs) are complex and new NEs are continuously increasing in significant numbers, leading to unknown-word issues in Biomedical Named Entity Recognition (BioNER) tasks. One reason why biomedical NEs are complex is that they have many variations with the interchangeability of Roman numbers and Latin characters, spaces and hyphens, etc. The number of new biomedical research papers is increasing, wherein approximately two papers per minute, re- sulting in more than 1 million papers each year, are added to the PubMed database (Landhuis, 2016). With this number of publications, new NEs are constantly being reported. In the last few years, NER using pre-trained lan- guage models (LMs), such as BERT (Devlin et al., 2018), ELMo (Peters et al., 2018), and Flair (Ak- bik et al., 2019), has shown state-of-the-art perfor- mance. In the biomedical domain, pre-trained LMs such as BioBERT (Lee et al., 2019a) and BioELMo (Jin et al., 2019), which are BERT and ELMo trained on a biomedical domain text, have achieved the state-of-the-art performance in many biomed- ical natural language processing tasks including NER. However, only using previously trained LMs cannot cover the continuously increasing new enti- ties due to complex characteristics of biomedical NEs, lead to unknown words problem. Despite being used as approaches to avoid unknown words problem, subword segmentation (Sennrich et al., 2015; Kudo and Richardson, 2018) methods con- sider subwords represented as unique IDs, but not words or their synonyms. therefore, it is difficult for subword or character based LMs to cover biomedi- cal NEs , which are complex and contain various of expression described in section 3. Moreover, LM pre-training is costly, time-consuming, and computationally expensive. Training BioBERT on biomedical corpora based on the BERT model re- quires 10 to 23 days on eight NVIDIA V100 GPUs (Lee et al., 2019a). To deal with the complex and continuously in- creasing entities, the use of dictionary-based ap- proaches can be an effective …

Content truncated to 500 words