Research On Medical Report Generation Method Integrating Copy Mechanism And Representation Enhancement

Intelligent medical is a new health care service mode integrating artificial intelligence technology.The automatic generation of medical reports is an important task in the field of intelligent medical,which has been widely concerned by academia and business in recent years.This task generates a semi-structured medical report containing multiple sub-parts,such as chief complaint and current medical history,based on the patient’s self-report and doctor-patient dialogue in the consultation.The realization of this task can reduce doctors’ burden on write medical reports and effectively alleviate the strain on medical resources.Based on the characteristics of the medical report generation task,the following main work has been carried out:(1)The medical report generation method combining extraction and generation is proposed.Based on the feature that the medical report contains multiple sub-parts,the method of extraction and generation of molecular parts is adopted.First,the key sentences are extracted from the doctor-patient dialogue for each sub-part to reduce the interference of irrelevant information,and then the generation is carried out,which is then combined into a complete medical report.The experimental results on the CCL 2021 data set show that the performance of this method is 3.67%-8.99% higher than that of the main baseline model.(2)The medical report generation incorporating copy mechanism is proposed.A combination model of extraction and generation incorporating the Bio Copy mechanism is adopted.For the characteristics of medical reports that contain a large number of medical terms directly derived from doctor-patient dialogue,we had added a Bio Copy mechanism based on a combined model of extraction and abstraction,the Bio Copy mechanism added can effectively copy the medical terms in the key sentences to ensure the accuracy of the results.Relevant experiments show that the model has achieved better results than the main baseline model,with an increase of 2.61% and 0.59% respectively compared with the two baselines with the highest score(3)The medical report generation with representation enhancement is proposed.In order to enhance the text representation of the model,We adopted two methods in this paper: 1)The method based on prompt learning effectively uses the large-scale knowledge of the BERT layer to avoid the deviation of the sentence vector caused by the influence of word frequency and other factors.We constructed a prompt template containing [MASK]and combines with the sentence,then passed through a mask language model to obtain a vector representation of [MASK],which was used as a sentence vector.2)Based on the bifurcation regularization multi-headed self-attention mechanism,the model word vector focuses on semantic information from different angles.This method regularizes the attention matrix or output of multi-head self-attention,so as to constrain each attention head to make the features obtained from each other as orthogonal as possible,so as to capture more comprehensive semantic information.The experimental results on relevant data sets show that the effect of this model is better than that of all baseline models.To sum up,according to the characteristics of automatic medical report generation tasks,we had studied a method that incorporate copy mechanisms and representation enhancement,and achieved certain results.In the future,we need to conduct deeper research in aspects such as text representation.