Extracting biomedical entity relations from the literature using a recursive neural network

다국어 초록 (Multilingual Abstract)

Due to the trendy word “Big Data”, it is widely known that a large volume of data has inherent values. Therefore, it is becoming essential to analyze data in many areas ranging from social media to scientific literature. Among the various fields, the bioinformatics is considered as an important emerging field as to technological advancement is crucial and the domain expert is limited in number. The bioinformatics is an interdisciplinary field that combines data mining and biomedical domain, and it has two purposes. One is to extract patterns from data to gain insight that is hard to detect. The other is to reduce the manual effort in laborious task of analyzing data.
Understanding the relationship of entities in the biomedical field is a vital issue. Especially, it is necessary to extract entity relationships to aim for precision medical care that provides customized healthcare services. The main approach of this thesis is based on the Recursive Neural Network approach. The Recursive Neural Network model receives inputs in a form of a parsed tree. The parsed tree is a grammatical structure of a given sentence, so that the model can learn the sentence structure understood by humans. This thesis deals with two studies that focus on the relationship between two important words in a biomedical paper. The first study extracts possible adverse effects in drug combinations, and the second study extracts relationships such as activator / inhibitor in the chemical-protein relationship.

목차 (Table of Contents)

• 1 Introduction 1
• 2 Preliminaries: Recursive Neural Network 3
• 2.1 Text Mining 3
• 2.1.1 Recursive Neural Network and Parsed Tree 4
• 2.1.2 An Overview of Recursive Neural Network Models 4
• 2.1.3 A brief comparison between the Recurrent and the Recursive Neural
• Networks 7
• 2.2 Future Directions 9
• 3 Drug drug interaction extraction from the literature using a recursive
• neural network 11
• 3.1 Background and Problem Definition 11
• 3.2 Materials and methods 13
• 3.2.1 Model development 13
• 3.2.2 Preprocessing 15
• 3.2.3 Parsing sentences 17
• 3.2.4 Word embedding 18
• 3.2.5 Recursive neural network with treeLSTM 18
• 3.3 Evaluation Results 23
• 3.3.1 Experimental settings 23
• 3.3.2 DDI13 data 24
• 3.3.3 PK DDI data 28
• 3.4 Discussion 32
• 3.4.1 Robustness of our model 32
• 3.4.2 Regularization analysis 32
• 4 Chemicalgene relation extraction challenge results using recursive neural
• network models 34
• 4.1 Background and Problem Definition 34
• 4.2 Materials and methods 36
• 4.2.1 Pre-processing 38
• 4.2.2 Parsing sentences 39
• 4.2.3 Subtree containment feature generation 39
• 4.2.4 Position feature generation 39
• 4.2.5 Word Embedding 41
• 4.2.6 Recursive neural network with tree-LSTM 42
• 4.2.7 Ensemble method 44
• 4.2.8 Implementation detail 44
• 4.3 Post-challenge enhancements 45
• 4.3.1 Additional pre-processing 45
• 4.3.2 SPINN with additional pre-processing 45
• 4.4 Evaluation Results 47
• 4.4.1 Data corpus 47
• 4.4.2 Hyperparameter 47
• 4.4.3 Performance 49
• 4.4.4 Summary and future directions 51
• 5 Discussion and Conclusion 53
• 5.1 Discussion 53
• 5.1.1 SPINN model for the Drug-Drug Interaction Extraction Task 53
• 5.1.2 Error analysis from the Drug-Drug Interaction Extraction Task 55
• 5.1.3 Error analysis from the Chemical-Protein Interaction Extraction Task 58
• 5.2 Conclusion 59
• Bibliography 61