Research And Implementation Of Few-Shot Question-Answering Model With Model Compression

With the advent of the era of Internet information explosion,the amount of data added to the Internet every day far exceeds the ability of ordinary people to process information,and in the past,people used search engines to use keyword matching to help them filter information on the Internet.By taking the question and answer generated by people in natural life as an example,deep neural networks can learn the ability to use natural language to answer natural language,and with the power of computers,researchers have developed question answering systems to understand the questions raised by humans,quickly screen out information from the network,summarize and answer.As an intelligent information screening tool,the question answering system has broad prospects,but the bottlenecks of the current question answering system in practical application are:first,when actually training the question answering model,it is difficult to obtain a large number of labeled training data;Second,the question answering system using pre-trained language models occupies a large space and the reasoning speed is slow.In response to the above issues,this article proposes corresponding improvement methods.In the retrieval task stage,this article proposes a retrieval tool training algorithm-PSEQA based on prompt learning.This algorithm is based on a fill in the blank prompt learning template,using different soft prompts for different data domains and sentence types,enabling the model to purposefully learn general knowledge of retrieval tasks and distinguish unique knowledge within the data domain,enhancing the performance of the retrieval model in few-shot scenarios.In the reading comprehension task stage,based on the use of the blank filling prompt learning template,the output information of the searcher is fused in the template,and a PRFR reader algorithm is proposed.This algorithm helps the model to grasp the overall characteristics of the input sequence,and enhance the model’s reading comprehension ability and transfer learning ability.The loss of reader importance target was proposed,and the attention score of the reader model was used to guide the distillation process of the retrieval model.The two-stage distillation was connected in series,and a joint distillation method for the two-stage question and answer model was proposed.Build an actual Q&A system simulation platform,use JavaScript and Flask as the architecture to build web services,apply databases such as Faiss and MySQL to organize data,integrate the proposed retrieval and reader algorithms,and use a joint distillation compression model to achieve fast response to the Q&A system on the ground.The comparative experimental results of the proposed retrieval and reader algorithms demonstrate that PSEQA and PRFR perform significantly in small sample and zero sample scenarios.The effectiveness experiments conducted on the combined distillation method have demonstrated the effectiveness of the distillation method in reducing the spatial cost of the model.