Extracting drug–drug interaction (DDI) in the text is the process of identifying how two target drugs in a given sentence interact. Previous methods, which were limited to conventional machine learning techniques, we are susceptible to issues such as “vocabulary gap” and unattainable automation processes in feature extraction. Inspired by deep learning in natural language preprocessing, we addressed the aforementioned problems based on dynamic transfer matrix and memory networks. A TM-RNN method is proposed by adding the transfer weight matrix in multilayer bidirectional LSTM to improve robustness and introduce a memory network for feature fusion. We evaluated the TM-RNN model on the DDIExtraction 2013 Task. The proposed model achieved an overall F-score of 72.43, which outperforms the latest methods based on support vector machine and other neural networks. Meanwhile, the experimental results also indicated that the proposed model is more stable and less affected by negative samples.
Paper
Full text
Drug-Drug Interaction Extraction Based on Transfer Weight Matrix and Memory Network
Semantic Scholar · Computer Science · 2019
Abstract
Extracting drug–drug interaction (DDI) in the text is the process of identifying how two target drugs in a given sentence interact. Previous methods, which were limited to conventional machine learning techniques, we are susceptible to issues such as “vocabulary gap” and unattainable automation processes in feature extraction. Inspired by deep learning in natural language preprocessing, we addressed the aforementioned problems based on dynamic transfer matrix and memory networks. A TM-RNN method is proposed by adding the transfer weight matrix in multilayer bidirectional LSTM to improve robustness and introduce a memory network for feature fusion. We evaluated the TM-RNN model on the DDIExtraction 2013 Task. The proposed model achieved an overall F-score of 72.43, which outperforms the latest methods based on support vector machine and other neural networks. Meanwhile, the experimental results also indicated that the proposed model is more stable and less affected by negative samples.