The Research Of Biomedical Event Extraction Based On Combinational Learning And Self-training

With dramatic increasing in amount of biomedical literature, it becomes difficult for biomedical practitioners to efficiently access to the information which they are interested in for such a mass of biomedical literature. Therefore, it turns into a arrestive direction in biomedical information extraiction field that efficiently extracted managable and structured information from massive and unstructured text. Biomedical Event Extraction belongs to the scope of biomedical information extraction, and its objective is to extracte structured biological events information related to protein in unstructured text information on the molecular level.Machine learning methods have been widely utilized in biomedical event extraction research. In this paper, we also make use of machine learning approachs in biomedical event extraction, involving combination of learns, self-training model and kernel method. We separate event extraction into four common steps, including text pre-processing, trigger detection, argument detection and post-processing. In the phase of trigger detection, we combine different classifiers based on their decision theory and employe self-training approach to implement semi-supervised method and make use of unlabeled data. A trigger dictionary, which can decide whether a word is a candidate trigger, is established from the training corpus. And then features are extracted for ecah candidate triggers. After classification we determines whether a candidate trigger is a real trigger, and also assign an event type to each real trigger. In the argument detection stage, we construct trigger-protein pairs and draw on approach used in protein-protein interaction extraction to achieve the goal of argument detection. We divide argument dectction into simple detction and complex detection based on the definition of event types. In simple detection, we directly decide the whethe an argument belongs to a trigger. However, in comlex detection, we primarily judge whether a trigger-argument pairs exist a relationship. And if it exists, we use the same method which use in first stage to identify the exact type between these two entities in this pair. We use kernel method in this paper to recognise the relationship of trigger-argument pairs.We carry out our experiments on public corpus of BioNLP’09and BioNLP’ l1shared task. While unlabeled data corpus is selected from PubMed abstract document. Finally, result shows that our method can improve the performance of biomedical event extraction system and achieve a relative good extraction reslut.