Deep Learning Natural Language Generation System For Scientific Literature Based On Microservices

Natural language processing technology is widely used in today's Internet application scenarios such as product recommendation,public opinion analysis,and intelligent assistants.Among them,natural language generation has the characteristics of strong knowledge dependence and complex model structure,and its implementation is far more difficult than other models.To improve the performance of machine learning models on natural language generation tasks,this thesis explores several directions such as model structure,training methods,and data set production.The optimization method of the deployment of machine learning models in microservice scenarios is discussed,and the research is carried out with service traffic as the starting point.The main work and achievements of this thesis include the following three aspects:(1)Aiming at the problem of conditional natural language generation,this thesis proposes a subject-controllable text generation model.By modifying the mask matrix and normalization part of the Transformer model,a conditional text generation model and a two-stage decoding method are proposed.Through supervised training,the model can learn the labeling standards of the dataset,thereby generating sentence pairs that meet a certain relationship.The experimentally verified that the generated samples have high text quality and can play a role in data enhancement.(2)Aiming at the problem of multi-task text generation in the field of scientific literature,this thesis proposes a multi-task learning method based on structured data sources,so that a model can be used for various text generation tasks.In addition,this thesis proposes the first structured multi-task data set,the Chinese Scientific Literature data set,which can derive a large number of natural language generation tasks by using naturally annotated data,and has high data quality.After that,this thesis verified the proposed method on the T5 pre-training model,and found that it is capable of various natural language generation scenarios.It is currently the first Chinese multi-task natural language generation model.(3)Aiming at the problem of machine learning model deployment,this thesis explores the flow control scheme in the microservice scenario,and proposes an adaptive flow control algorithm based on operations research methods.Model the hardware usage and network traffic scenarios of microservice nodes,and calculate the current flow thresholds of each service in real time.After comparative experiments,this method of dynamically calculating the current limit threshold can increase the system throughput and resource utilization in various traffic scenarios compared to the traditional static method.