Research And Application Of Key Technologies Of Silk Knowledge Base Question Answering System Based On Deep Learning

Silk plays an important part of Chinese traditional culture.Fully researching and using silk cultural relics is of great significance to carry forward the spirit of traditional culture and enhance national self-confidence and influence.At present,some progress has been made in the construction of digital platforms for silk cultural relics.However,there is still a lack of related research and applications for deep excavation of silk cultural relics and the combination of current academic frontiers,which cannot meet the increasing cultural needs of people.With the continuous development of information technology,traditional keyword search services are increasingly difficult to meet people's needs.Question answering system,as an advanced information retrieval method,enables machines to understand natural language questions raised by humans,and is gradually becoming a new trend of human-computer interaction.In addition,the establishment and development of the knowledge base,on the one hand,provides luxuriantly data resources for the question answering system.On the other hand,it also puts forward new requirements for the performance and accuracy of the question answering system.Current knowledge-based question answering systems have some problems with low entity recognition accuracy and out-of-vocabulary issue.Aiming at dealing with the above problems,this article has conducted in-depth research and discussion on named entity recognition tasks and relation detection tasks.In this paper,we also design and implement a question answering system based on the silk knowledge base.The Main tasks as follows:1)In order to improve the accuracy of named entity recognition of the question answering system,and aiming at the characteristics of entity naming in the silk field,this paper first proposes a new entity classification method and word segmentation method.Then we propose a fusion model based on BERT and CRF.By using transfer learning method,first,we use pre-trained data of BERT to initialize the model.Then combined with specific named entity recognition tasks,the parameters are fine-tuned accordingly.And the results were parameter constrained and decoded through CRF layer.Finally,through experimental verification and comparison,the BERT+CRF fusion model achieved a 2.4% improvement in the accuracy of named entity recognition Task.2)To alleviate out-of-vocabulary problem,a new relation detection model based on adversarial learning adapter is proposed.First,the relation vector is pre-trained by Trans H model.Then through adversarial learning,we train a relational adapter which continuously confront the generator and discriminator,to learn the knowledge base relational representation.And then further relation features are extracted and strengthened through BERT to obtain the final relation embedding representation.Through comparative experiments,it is proved that the problem of the failure of the relation vector that appears in the test set but not in the training set has been well solved,which has largely alleviated the problem of lexical overflow.3)Based on the above two models,a question and answer system,based on the silk knowledge base,with more functional modules is established.The Q & A system is decomposed into front end and back end and structured by B / S architecture.it provides a more common front-end for mobile and web pages.Additionally,it also encapsulates the Q & A system into a back-end service,and provides an interface of question Answering services.