| The "Four Insurances and One Housing Fund" is an important part of Chinese social security system.The relevant regulations are derived from policy and regulatory documents,which have regional characteristics and mass characteristics.It is difficult for the people to understand the knowledge they need,which generate a large number of consulting needs that will exist for a long time.The modernization of government governance capabilities has been a hot topic in recent years.The research on the question answering system in the field of "Four Insurances and One Housing Fund" can not only allow the people to understand the provisions of policies and regulations at any time,but also reduce the government’s workload,which has important research significance.This thesis first researches the question intent classification technology.Recognizing the user’s question intent can reduce the knowledge retrieval scope of the question answering system and improve its response speed and accuracy.Since there is no question intent classification benchmark in the field of "Four Insurances and One Housing Fund",this thesis combines the relevant policies and regulations in the field of "Four Insurances and One Housing Fund" to manually analyze and summarize the question intent classification benchmarks from the perspective of business,build a dataset of question intent classification in the field of "Four Insurances and One Housing Fund".In order to solve the problem that the question features are sparse,and the neural network model is difficult to accurately understand the question semantics,this thesis proposed a question intent classification algorithm fused with word frequency information.The algorithm integrates word frequency information on the basis of the existing neural network model to make the model easier to understand the semantics of the words,thereby improving the classification ability of the model.This thesis compares the accuracy and F1 value of six neural network models before and after fusion of word frequency information on three classification datasets,which shows that the classification ability of the model after fusion of word frequency information has been effectively improved.This thesis finally studies the related technologies of question answering system in the field of "Four Insurances and One Fund",proposes a technical route which uses the "Four Insurances and One Housing Fund" field policy and regulation documents and a large number of Q&A pairs crawled from the Q&A community as the knowledge base.When getting the answer,this system retrieves the answer from knowledge base through the algorithm based on text similarity.The algorithm first obtains the Q&A pairs related to the user’s question through the text similarity calculation algorithm based on Word2 Vec,and then obtains the policy terms related to the answer of the Q&A pairs through the text similarity calculation algorithm based on TFIDF.After the final screening,the related Q&A pairs and related policy terms are returned as answers together.The question intent classification plays a role in narrowing the scope of knowledge retrieval in the process of obtaining answers,and it effectively improves the response speed of the question answering system after application.After implementing the question answering system according to this technical route,the actual use effect shows that the question answering system can answer the questions raised by the user very well. |
