Research On Question And Answer Technology Of Corporate Financial Audit Based On Deep Learning

Due to professionalism,a large knowledge system and rapid renewal of corporate financial audit,auditors may be confronted with various problems and situations in work.In order to solve these problems,they need to acquire professional knowledge effectively and correctly.However,most current softwares of auditing information only focuse on management of audit business,but offer no function relenting to acquisition and retrieval of auditing knowledge.Therefore,it is of profound research significance and high application value to develop the key technology of a question-answering system targeting the field of corporate financial audit,answer auditors' questions rapidly and correctly with natural language and offer them professional,intelligent and convenient question-answering service.In addition,on the basis of the theory of deep learning,this paper studied question-answering technology relenting to corporate financial audit.In this paper,the question-answering system was divided into two key modules,namely entity identification and question-answering retrieval.In respect of entity identification,a method of field entity identification combining a recurrent neural network model(Bi-LTSM)based on long-and-short-term memory with a model of conditional random fields(CRF)was created.In terms of question-answering retrieval,a model of question-answering retrieval based on multi-granularity text matching was established and a question-answering system in the field of corporate financial audit was constructed.Entity identification in the field of corporate financial audit contributed to understanding the core meaning of users' questions and its recognition effect could influence subsequent answer matching directly.At this stage,linguistic data were processed to form characteristics integrating field knowledge,such as a domain dictionary and demonstrative words.Then,entity identification was conducted by combining Bi-LTSM with CRF,which solved the problem of poor entity identification in the field of corporate financial audit.At the stage of question-answering retrieval,modeling of local information and overall information of sentences was conducted with a two-path coding model based on the attention mechanism.Then,an interactive text matching model was used for direct interaction on the input layer.A model of multi-layer convolutional neural network(CNN)was utilized to extract fine matching characteristics from the interaction matrix and introduce the stage of entity identification for identifying entities as input features,improving its effect.Meanwhile,the final matching scores were figured out by combining two matching approaches and the answer with the highest score was replied to the user,enhancing the accuracy of answers.Finally,this paper conducted a contrast experiment relenting to algorithms and models mentioned above and confirmed effectiveness and application value of the Bi-LSTM+CRF-based entity identification algorithm in the field of corporate financial audit and the question-answering retrieval model based on multi-granularity text matching.