Research And Implement On Domain Oriented Automatic Question Answering System

Automated question and answer techniques are designed to answer questions entered by the user in the form of natural language by first making semantic sense of the questions and then querying the area where the answers are located using automated retrieval techniques.Auto-questions are divided into open domain auto-questions and restricted domain autoquestions,the main difference between the two is whether the query questions and background knowledge base belong to the same domain.Research towards a domainrestricted auto-quiz.It involves natural language representation,information retrieval,semantic matching and other technologies,which have good research significance and commercial value.With the development of deep learning in various fields in recent years,the technology of restricted domain auto-questioning based on deep learning has also made great strides.Compared to traditional question-and-answer techniques,the deep learning-based approach avoids mechanical retrieval and offers a significant improvement in both predictive answer accuracy and question-answer efficiency.Nonetheless,there are still some problems: 1)there is a lack of attention problem when extracting semantic features,which cannot capture the semantic features of long texts;2)there is a problem of multisemantic context recognition when representing texts,which cannot identify the correct semantics of multisemantic words in different contexts;3)there is a problem of category imbalance when predicting answers,which cannot predict the correct answers to problems that occur less frequently.Based on the above issues,this paper proposes an enhanced self-attention question and answer model.For the problem of lack of attention,the model introduces multiple rounds of self-attentiveness mechanism,which enhances the influence of text-question and problemtext interaction through multiple rounds of self-attentiveness query on the input sequence;for the problem of multisemantic context recognition,the model characterizes the input words in multiple dimensions from both static and dynamic word vectors to improve the text context characterization ability and fully extract the grammatical and semantic features of the input text;for the problem of category imbalance,the model replaces the cross-entropy loss function with the focus loss function,while oversampling the input data and expanding the data volume through translation and synonym replacement,etc.,to reduce the influence of category imbalance.Based on the previous research work,this paper also designs and builds a domain-limited automated question and answer system that allows users to interact in real time to verify the performance of the proposed model.