An Insight Extraction System on BioMedical Literature with Deep Neural Networks

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2691–2701 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics An Insight Extraction System on BioMedical Literature with Deep Neural Networks Hua He1, Kris Ganjam2, Navendu Jain2, Jessica Lundin2, Ryen White2 and Jimmy Lin3 1 Department of Computer Science, University of Maryland College Park huah@cs.umd.edu 2 Microsoft {krisgan,navendu,jelundin,ryenw}@microsoft.com 3 Cheriton School of Computer Science, University of Waterloo jimmylin@uwaterloo.ca Abstract Mining biomedical text offers an oppor- tunity to automatically discover important facts and infer associations among them. As new scientific findings appear across a large collection of biomedical publica- tions, our aim is to tap into this literature to automate biomedical knowledge extrac- tion and identify important insights from them. Towards that goal, we develop a system with novel deep neural networks to extract insights on biomedical litera- ture. Evaluation shows our system is able to provide insights with competitive accu- racy of human acceptance and its relation extraction component outperforms previ- ous work. 1 Introduction Biomedical literature offers a rich set of knowl- edge sources to discover important facts and find associations among them. For instance, MED- LINE contains over 18 million references to arti- cles published since 1946 and sourced from over 5500 journals worldwide (Simpson and Demner- Fushman, 2012). Two major processing tasks performed on the biomedical text are: (1) iden- tify and classify biomedical entities (NER) into predefined categories such as proteins, genes, or diseases, and (2) infer pair-wise relationships among named entities e.g., protein-protein interac- tion (Poon et al., 2014), gene-protein, and medical problem-treatment. This paper presents a system that processes biomedical text to extract two specific types of re- lationships among biomedical entities: (a) cause- effect and (b) correlation. This system is motivated by the need to bet- ter automate biomedical knowledge extraction and identify important information from them, as new scientific findings appear across a large collection of publications. For instance, given user sleep pat- terns, existing biomedical research can be better utilized to provide insights: inform about poten- tial effect (e.g., “diabetes”, “obesity”) due to the cause (e.g., “sleep disorder”) and suggest appro- priate treatment. Since biomedical articles usually have title and abstract summarizing the contents of the full-text article, we focus on extracting the two relationship types from them. Unfortunately, mining this sum- mary data still poses several key challenges. Sim- ilar to full-text, this data comprises unstructured text with domain-specific vocabulary, issues of synonymy (e.g., “heart attack” vs. “myocardial infarction”), acronyms, abbreviations and rapidly evolving terminology due to new scientific discov- eries. While the titles are short and informative, they do not contain the key information that would be contained in the abstract. Many of these challenges are also applicable for biomedical relation extraction. Further, iden- tifying particular relation types is challenging be- cause relations are expressed as discontinuous spans of text , and the relation types are typically application-specific. Finally, there is often little consensus on how to best annotate relation types resulting in lack of …

Content truncated to 500 words