The Study On Joint Models For Biomedical Event Extraction

With the application of high throughput sequencing technology in the field of the biomedical science, more and more biomedical experimental results have been published in text so that the amount of the biomedical literature is increasing dramatically. But the information in the text is hard to be used by the researchers because it is unstructual. The research of biomedical text mining employs all kinds of technologies, such as natural language processing, biomedical informatics, computational linguistics, and artificial intelligence etc. to extract the semantic relations among the biomedical entities to construct the knowledge base which is structural and easy to use. The common tasks of biomedical text mining include Information Retrieval (IR), Information Extraction (IE), knowledge base construction, and knowledge discovery etc. Among them IE task includes Named Entity Recognition (NER), Named Entity Normalization, and entity relation extraction.In recent years, biomedical event extraction became a popular research point in information extraction. It originated from the shared task sponsored by Tsujii laboratory at university of Tokyo in 2009. The Bio-event extraction aims to extract the nested events related to protein entities in the text. The organizer published the general and consistent definition for bio-event and provided related data set with estimation criteria. In the data set, the protein entity names have been annotated. The participants were required to detect the triggers which show the changes of the protein activities or status. Furthermore, the nested semantic relations among the proteins and triggers should be extracted. Because the task is very complex, the performance of the published the bio-event extraction systems can't reach the standard of application yet. So it needs more researches on this point. In this paper we focused on the joint models to extract the biomedical events. Before that we studied the method of the trigger detection and finished the pipelined method of the bio-event extraction. The main researches summarized as follows:(1) The model of sequence tagging was applied to detect the triggers. The event extraction from biomedical literature plays important role in the knowledge mining in biomedical domain. The trigger detection is the key step in biomedical event extraction. In this paper we treated the trigger detection as the task of sequence tagging and used the CRF model to attain it. To train the model, we used rich features including lemma, context, phrase label word cluster and learned trigger dictionary to build several kinds of CRF models based the word. Then we choose the best model for each type of triggers to combine a hybrid model. The evaluation on the BioNLP-ST 2009 data set shows that our approach achieved good performance, which laid foundation for biomedical event extraction. In addition, the similar strategy was applied to the task of ChemistryNer in BioCreative?. The performance of the subtask of CDI and CEM got the first and the second position in the challenge respectively.(2) A pipeline-based method was used to extract the bio-events. The pipeline-based strategy is one of the mainstream methods of event extraction which has advantages and inevitable disadvantages. Accordingly we attempted to extract the events in line with the pipeline flow before studying on the joint models. The approach contains three sub steps as following:firstly, based on the sentence parse we extracted the target dependency sequence from which we obtained the candidate event pairs; fourthly, the candidate event pairs were classified by the two-stage classification. The first classifier partitioned the pairs into 9 classes according to the 9 event types. The second classifier identified the each type of instances positive or negative; in the end, a post-processing step was used to construct the events. When evaluated on the BioNLP-ST2013 dataset, the precision overwhelmed all the participants in BioNLP-ST2013 but the overall performance was moderate.(3) The joint method based on tagging the entity chains was proposed to extract the events. We proposed a novel and efficient approach for extracting nested biomedical events jointly. Based on this method, the trigger detection and event edge extraction were implemented at the same time. We decomposed the bio-events into entity chains from the protein argument to trigger on the highest layer of nested event in concept. The entity relations of bio-event are semantic, so they can be represented by the dependency parsing. Therefore, based on dependency parsing, we extracted the target sequences that contained biomedical entity (trigger/argument) chains firstly. Secondly, the Condition Random Fields (CRFs) model was used to tag the entity chains. After tagging, a strategy was used to optimize the results. Finally, the post-processing step was used to construct the event structures. The experiment results showed that we got the performance of 47.3 in F-score which is promising when compared with the joint ML-based system in BioNLP-ST2013. Furthermore, we estimated the results of the trigger detection. The F-score reached 68.03 on BioNLP-ST2009 corpus and 71.33% on BioNLP-ST2013.(4) The joint model based on the structure prediction strategy with inexact search was studied to extract biomedical events in this paper. This strategy treated the output of the event in the sentence as structure which was constructed incrementally. In the experiment the structure perceptron with Beam search decoding was adopted to train the model. In decoding, the complexity of the task is sure to make the search space so huge that exact inference is intractable. Therefore beam-search along with early-update strategy was employed to perform inexact decoding which make the joint model feasible. The advantages of this model rest on that both the local features and the global features can be used. We perform the experiments on BioNLP-ST2013 dataset and get a medium performance with the F-score 43.8%. Compared with the joint model based on tagging the entity chains in the previous, this strategy showed advantages in the events with multiple arguments though it was moderate on the overall performance. Thus structure prediction model showed positive contribution on the extraction of complex biomedical events.