| Question answering can return correct answer according to questions of users,which is a hot topic in the field of natural language processing.So far,the open-domain question answering has reached a high accuracy for simple questions,but the performance of biomedical question answering still needs to be improved.There are some problems in biomedical question answering,such as complex relationship between biomedical entities,difficulty in understanding the semantics of question and answer pairs and so on.To solve these problems,this thesis studies the key technologies of biomedical question answering system,such as the semantic relationship between question and answer pairs and the semantic understanding based on professional knowledge.Main research contents are list as follows:（1）Biomedical question answering based on semantic mining of question and answer pairs.Biomedical question answering systems need to capture the semantic connections from complex question and answer pairs.In order to fully capture the semantic connections,this thesis proposes a multi-unit co-attention mechanism.Firstly,the co-attention mechanism is adopted to jointly reasoning question attention and answer attention,in which answer represention is used to guide question attention and question representation is used to guide answer attention,and obtain the semantic connections of question answer pairs.Then,several parallel co-attention mechanism are introduced to promote the semantic representation of question and answer pairs through different co-attention units.Experiments on MEDIQA dataset show that the multi-unit co-attention mechanism proposed in this thesis can deeply mine the semantic connections of question and answer pairs,and effectively improve the classification performance of biomedical question answering model.（2）Biomedical question answering based on knowledge selection.Biomedical text has many entities and complex entity relationships.Sometimes it is impossible to fully capture the semantics according to the context,and some professional knowledge is needed.In order to introduce knowledge in biomedical domain,this thesis first constructs a knowledge base based on question answering tasks.Then,the knowledge representation learning method is used to obtain the representation of entities and relationships.Next,we attend knowledge using the semantic representation of question and answer respectively,to obtain the knowledge representation of questions and answers.Finally,the gating mechanism is used to integrate knowledge representation and semantic representation to obtain the knowledge and semantic joint representation of questions and answers respectively,which is used for classification of question and answer pairs.Experiments on MEDIQA dataset show that the biomedical question answering model based on domain knowledge selection can select useful knowledge to assist the classification of question answering pairs,and improve the performance of biomedical question answering model.（3）Biomedical question answering based on generation of adversarial samples.Type information can play an important role in assisting question and answer pairs classification.In order to accurately capture the type information,this thesis adds perturbation of type information in question to generate adversarial samples.Then,the adversarial samples are added to the training set.The biomedical question answering model based on semantic mining of question and answer pairs and the biomedical question answering model based on knowledge selection are trained separately to enhance the ability of model to capture type information.Finally,we use collaborative learning to make these two models learn from each other,and promote the fusion of domain knowledge and semantic information.Experiments on MEDIQA dataset show that biomedical question answering based on generation of adversarial samples can accurately capture the type information of questions and answers,and collaborative learning can effectively integrate the semantic information and knowledge information of different models to further improve the performance of model.Researches of this thesis provide effective methods to improve the performance of biomedical question answering.This method can also be used to other question answering tasks,which has certain universality. |
