Research Of Antibiotic Resistance Event Extraction Based On Attention Mechanism And Contrastive Learning

Antibiotic resistance is becoming a major problem due to human abuse and overuse of antibiotics,which has sparked the emergence of antibiotic-resistant bacteria.For clinical treatment and drug development,a thorough understanding of the mechanisms of antibiotic resistance is essential.Since the abundance of research on antibiotic resistance mechanisms in biomedical literature,effectively extracting structured information from the text has become an important research direction.Existing work models the extraction of structured information on antibiotic resistance as a biomedical event extraction task,using natural language processing techniques to achieve automated information extraction.However,existing methods have not produced ideal results in the task of extracting antibiotic resistance events due to the complexity of the structure of antibiotic resistance events and the limited amount of annotated data.To address these issues,this thesis uses attention mechanism to model event structures and employs contrastive learning to achieve data augmentation and learn features of different event types,effectively completing the task of extracting antibiotic resistance events.The following are the primary contributions of this thesis:(1)This thesis suggests an attention-based cascade model for extracting antibiotic resistance events to solve the issue of complicated and challenging modeling of biological event structures.Three cascaded decoders and a shared text encoder make up the model.The proposed model breaks down the event extraction work into a number of interconnected subtasks,including as event type recognition,trigger word extraction,and element extraction,enabling the identification of complicated event structures at various stages.Attention mechanism is used to model text semantics and event structures,thereby addressing the difficulty of modeling antibiotic resistance events.The effectiveness of the proposed model has been tested using experiments on common biomedical event extraction datasets MLEE and antibiotic resistance event extraction ABEE.(2)In this thesis,a contrastive-learning-based antibiotic resistance event extraction model is suggested to solve the issues of insufficient antibiotic resistance event annotation data and uneven distribution of event types.By using random dropout and trigger word replacement,respectively,the model creates an unsupervised contrastive learning task and an event-based supervised contrastive learning task.It then learns the features of various event types by pre-training text representation modules to further enhance the model’s capacity to learn event structures.The proposed model incorporates external knowledge embeddings to enrich the semantic information of word representations in order to more accurately model biomedical things.Experimental evidence demonstrates that the method suggested in the thesis efficiently addresses the issues of insufficiently annotated data and uneven distribution of event types,and improves the effectiveness of the model on the antibiotic resistance event extraction task.In order to automatically extract structured data related to antibiotic resistance events from biomedical literature,this thesis proposes an event extraction method based on attention mechanism and contrast learning.The proposed model increases the effectiveness of knowledge acquisition and furthers the study of antibiotic resistance mechanisms.