Research On Entailment Recognition Of Biomedical Questions Based On Multi-tasking

With the popularity of computer technology and the rapid development of the biologicalfield,the growth of the biomedical literature is exponential.How to quickly and efficiently obtain biomedical knowledge from massive data involves information extraction technology.As an important task of information extraction,recognition of question entailment aims to recognition the implication relationship between pairs of question texts.In the field of biomedicine,question entailment recognition has important applications in information retrieval and question-answering systems.This thesis focuses on the recognition of implication relationships for the pairs of biomedical question.The main research contents are as follows:Research on biomedical questions entailment recognition based on multi-tasking.There are a large number of question-answering task corpora in the field of biomedicine,and the rational use of these corpora can effectively assist biomedical questions with recognition performance.First,the shared Bio BERT model is used to learn to obtain question-answering sentence pairs and contextual word representations of question-entailemnt sentence pairs.Then,construct the semantic representation of question-answering sentence pair and question-entailment sentence pair based on multi-tasking,which are used to classify two tasks.In-depth exploration of the impact of different neural network models on the learning performance of the question representation.Experiments show that multi-task learning can effectively improve the recognition performance of biomedical questions through sharing with biomedical question-answering task representation.Research on biomedical questions entailment recognition based on the characteristics of question types.The questions of the same type has an implication relationship,and the characteristics of the question type have an auxiliary role in the recognition of the question entailment.First,based on the Bio BERT model training,a teacher model of question type classification is obtained.Then use the predicted probability distribution of the teacher model to guide the classification of the student model based on the question type template,so that the student model can obtain the pre-training language knowledge of the teacher model and the template knowledge of the student model itself.Finally,the question type representation is obtained based on the student model,which assists the recognition of the question entailentbased on multi-tasking.Experiments show that the teacher-student model can integrate the pre-training language knowledge of the teacher model and the template knowledge of the student model,which can effectively improve the recognition performance of biomedical questions.Research on biomedical questions entailment recognition based on single-generation long and short-term knowledge distillation.The single-generation knowledge distillation can use the knowledge learned early in the model itself to guide the later learning of the model with high quality.In the process of question entailment recognition model training,the question entailment recognition model is obtained based on each epoch training as a short-term teacher model,and the predicted probability distribution is used to guide the learning of the next epoch.At the same time,multiple consecutive epochs are aggregated into disjoint mini-generation.Based on each mini-generation training,the question entailment recognition model is obtained as a long-term teacher model,and the predicted probability distribution is used to guide the next mini-generation.Experiments show that single generation long and short-term knowledge distillation learning strategy can effectively use the prior knowledge learned by the model to guide the model's own learning,and further improve the recognition performance of biomedical questions.In this thesis,based on the realization of a higher biomedical question entailment recognition performance,it can also be extended to the text entailment recognition task in the general field,with domain universality.