Research On Trigger Detection In Biomedical Event Extraction

Relation extraction has shifted to complex relation extraction from simple binary relation extraction in recent years. And biomedical event extraction has been the research focus. Trigger, argument and the event type need to be detected when extracting an event. Event extraction systems consist of at least two parts:trigger detection and argument detection, while trigger detection is the preceding task. Thus, trigger detection is of great importance in biomedical event extraction.Two kinds of methods on trigger detection are described in this work:hybrid method and deep learning method based on Stacked Denoising Autoencoders (SdA). In hybrid methods, multiple single classifiers are constructed firstly based on rich manual features including dependency and syntactic parsed result. Then multiple predicting results are integrated by set operation, voting and stacking method. Deep learning method based on SdA, adopts the word embeddings of words in linear window of the candidate triggers rather than rich manually designed features and reduces the dependence on biological expertise. And we preliminary prove the effectiveness of feature learning.Experiments will be conducted on the public evaluation corpus BioNLP’09 and BioNLP’11. Hybrid method will be evaluated on the two corpora and deep learning method based on SdA will be evaluated on BioNLP’09. The experimental results show that hybrid methods outperform single classifier and stacking method exceeds set operation and voting. The F-scores on BioNLP’09 and BioNLP’11 are 73.79% and 74.25% respectively. Under the condition of not adopting a lot of artificial features, deep learning method based on SdA achieves an F-score of 57.04% and proves the effectiveness of feature representation. On the base of word feature, adding the learnt feature from SdA, F-score achieves 74.41%, which reduces the artificial features and outperforms all artificial features on PA by 1.41%.