|
Biomedical event trigger extraction, as one of the sub-task of biomedical event extraction, plays an important role in biomedical research. A biomedical event trigger is a word or phrase that marks the emergence of a certain biomedical event. The recent works are usually based on the rule and the machine learning methods. However, the rule-based methods heavily rely on the concrete rules enumerated by the field expert and usually need a large amount of expert knowledge. The machine learning-based approaches usually utilize many handcraft features such as n-gram, lexicon, pose-tag and shortest dependency path. As a result, these methods based on machine learning can suffer from handcraft engineering with expensive time costs and the problem of generalization in the field transition. With the popularization of deep learning techniques, some effective frameworks in Natural Language Processing (NLP), such as adversarial training, self-attention mechanism, graph convolutional network, have been proposed to enhance the model performance for the NLP, especially the information extraction. As a task in the information extraction field, the frameworks mentioned above have been applied in the biomedical trigger identification subtask. This paper attempts to employ an external version of the recurrent neural network (RNN), i.e., bidirectional Gated Recurrent Unit (Bi-GRU) network, to extract the biomedical event trigger existing in the biological literature. Specifically, we first transform each token and entity label in the sentence to a word sequence with token index and an entity label sequence with entity label index. Subsequently, the above two sequences will be fed into the embedding layer to obtain the concatenated tensor between them. Moreover, we put the tensor into the Bi-GRU to generate the contextual encoding, which will be fed in a linear layer with an activation function to predict the probability distribution of the trigger. The final experiment on the MLEE dataset confirms that the proposed model can achieve comparable performance with an F-score of 78.82%.
|
