2016 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)
Biomedical Event Extraction Based on Distributed
Representation and Deep Learning Anran Wang, Jian Wang*, Hongfei Lin, Jianhai Zhang, Zhihao Yang, Kan Xu School of Computer Science and Technology, Dalian University of Technology, Dalian, China E-mail: * wangjian@dlut.edu.cn
Abstract—The two main problems of biomedical event B. Event Argument Detection
extraction are trigger identification and argument detection which Event argument detection is the step for examining whether can both be considered as classification problems. In this paper, entities have relation with the triggers and detecting the relation we propose a distributed representation method, which combines context, consisted by dependency-based word embedding, and type. We converted the word embeddings of the words on the task-based features represented in a distributed way on deep dependency-path between trigger and entity to a matrix as learning models to realize biomedical event extraction. The representation of the logical semantic relations. Then we experimental results on Multi-Level Event Extraction (MLEE) concatenated the matrix with their POS and relative distance corpus show higher F-scores compared to the state-of-the-art SVM vectors. We also added the type of trigger and entities. Finally, method. This demonstrates that our proposed method is effective we trained Convolutional neural network (CNN) with the input for biomedical event extraction. layer formed by these distributed representations.
Keywords—Biomedical event extraction; Distributed III. EXPERIMENT
representation; Deep learning; Convolutional neural network We evaluated our proposed method on MLEE corpus [4], and it achieved a better F-score compare to the baseline on I. INTRODUCTION trigger identification and event extraction. This indicated our A biomedical event is normally described by event type, method’s efficiency. We also conducted feature analysis tests trigger and arguments. The goal of extraction is to detect and and the results showed that task-based features contributed to the classify the type of trigger and the event arguments. Due to rich performance. feature engineering, support machine vector (SVM) previously has reached the best performance [1]. However, massive feature IV. CONCLUSION engineering and semantic gap from SVM created obstacles In this paper, we have presented a distributed representation when exploring more semantic features. Meanwhile, the deep method that combines dependent context formed by word learning methods have achieved considerable progress on some embedding with task-based features of biomedical text. Then we natural language processing tasks such as sentence fed it to deep learning models to realize biomedical event classification[2]. In this paper, we proposed a distributed extraction. This method avoided the problems of semantic gap representation method on deep learning models. Our goal is to and dimension disaster from traditional one-hot representation realize event trigger identification and event argument detection methods and achieved promising results with few manually to complete biomedical event extraction. extracted features, instead of complex feature engineering. II. METHOD ACKNOWLEDGMENT Our main work proceeded as follows. First, we trained This work is supported by the grants from the National Key dependency-based word embedding from all available PubMed Technology R&D Program (2015BAF20B02), Natural Science abstracts. Then, we formed the distributed representation for Foundation of China (61572098, 61572102, and 61602078), and deep learning models for the two subtasks. Liaoning Province Natural Science Foundation of China A. Event Trigger Identification (2014020003). The aim of trigger identification is to predict the trigger type REFERENCES of words. First, we concatenated word embedding trained [1] D. Zhou, D. Zhong, “A semi-supervised learning framework for previously for words in the window size around candidate biomedical event extraction based on hidden topics,” Artificial triggers. Then, we multiplied the topic distribution of every word intelligence in medicine, vol. 64(1), pp. 51-58, 2015. acquired by Latent Dirichlet Allocation (LDA)[3] to represent [2] M. Ma, L. Huang, B. Xiang, B. Zhou, “Dependency-based Convolutional the approximate sentence topic distribution. Third, we randomly Neural Networks for Sentence Embedding,” Meeting of the Association initialized two vector tables for the POS of words and the for Computational Linguistics and the International Joint Conference on Natural Language Processing, 2015. distance of words to closest entity in the dependency tree. [3] D.M. Blei,A.Y. Ng,M.I. Jordan, “Latent dirichlet allocation,” Journal Finally, the concatenation of them formed the representation of of Machine Learning Research, vol. 3, pp. 993-1022, 2003. candidate triggers and was fed to the deep neural network. A [4] S. Pyysalo, T. Ohta, M Miwa, H.C. Cho, J.I. Tsujii, et al., "Event softmax layer was used to realize classification. extraction across multiple levels of biological organization," Bioinformatics, vol. 28(18), pp. i575-i581, 2012.