Research And Implementation Of Chinese Intelligent Question Answering Technology Based On Machine Reading Comprehension

The intelligent question answering technology can help people to accurately find the required content in a variety of network information,effectively solving the problem of accurately positioning the answer in the traditional retrieval process,and machine reading comprehension is a new method of the intelligent question answering technology without the need to establish Large-scale question and answer template library,thereby greatly reducing the cost of building,maintaining and operating the template library.This paper explores the problem of Chinese reading comprehension based on the transfer learning model,optimizes the structure of the current cutting-edge transfer learning model,and builds a Chinese machine reading comprehension model with excellent performance.The specific research content is as follows:(1)This paper proposes a method of feature set construction based on multiple text mining methods to help the model simulate people's usual reading habits.There are three main types of prior features that are excavated: context prior features,query prior features,and context-query interaction features.Among them,the context's prior features include the features by part-of-speech tagging and named entity recognition.The query's prior features include the features by part-of-speech tagging,named entity recognition,and category.Context-query interaction features include co-occurrence features and distance features.Because of the limitation of transfer learning on the length of the document,a set of multi-document,multi-answer training set,label set construction methods and sample equalization strategies suitable for transfer learning models are shown in detail.(2)This paper proposes a transfer learning model called BERT-wwm-MLFA based on multi-layer full attention.The unidirectional and bidirectional attention mechanisms are combined into a full attention mechanism based on the fusion mechanism.According to the different semantic extraction capabilities of different layers in the benchmark model BERT-wwm,the full attention mechanism is used to interact and stitch the encoding results of different layers of the benchmark model,And then use the pointer network to output the results.Finally,an improved multi-answer extraction module and corresponding loss function are designed for the case where the answers come from different fragments but the same document.The adversarial training method is used to improve the robustness and accuracy of the model during training by adding disturbance factors.The experiment proves that the BERT-wwm-MLFA model proposed in this paper can achieve better prediction results,reaching ROUGE-L of 0.837 and BLEU-4 of 0.738.(3)In order to achieve the preliminary application of the machine reading comprehension model in intelligent question answering,this article uses the Chinese machine reading comprehension model in the military field and the ES search engine to build a Chinese intelligent question answering system,which supports user login,automatic question answering,and related Document export and other functions.Besides,It also supports the expansion of the Q&A model functions in other fields,providing possibilities for intelligent Q&A in multiple fields.