Design And Implementation Of Financial Question Answering System Based On Knowledge Representation

In recent years,with the rapid development of network,the promotion of 5G and the popularity of mobile devices,a variety of information appears in front of people.Although this has brought some convenience to people’s life and work,the massive amount of information has also caused some trouble.Many homogeneous and useless information makes it difficult for people to find the information they want when using search engines.It is necessary to carefully screen the useful information,especially for specific fields.With the development of information extraction,big data and other related technologies,it provides a good basic condition for the generation and optimization of automatic question answering system.The methods used in question answering system can be roughly divided into two kinds: semantic analysis and information retrieval.Semantic analysis aims to learn and understand a text through some natural language processing means,express it into some logical expression,query the knowledge base according to the logical expression,and finally obtain the answer;Information retrieval aims to search the relevant candidate answers from the database through some methods,and sort them through the comparison of some characteristics of questions and questions to get the final answer.Based on the project of a financial company and the financial field,this paper studies the relevant technologies and constructs a question and answer system based on the financial field.Compared with the general field,there are few ready-made data sets in the financial field,and some data have strong timeliness,so the collection and sorting of data is a very important part.The financial question answering system is based on the Bert pre-training model based on sub attention mechanism developed by Google and the improved twin network model,which aims to make the model better match questions and answers and improve the overall effect of the model.This model will first identify the named entity of the question,search the relevant answers in the database as candidate answers after error correction,and then train the question and candidate answers into vectors for text matching and scoring.The one with the highest score is the final answer.If this model encounters relevant entities that cannot be found in the database,it will conduct Q &a through a joint extraction model.If none,it will conduct chat through the accessed Baidu chat excuse.The experimental results show that compared with other benchmark models,the model proposed in this paper has achieved the best results,which reflects the effectiveness of this model.