| After rapid development of ten years, China Union Pay has been covering the country's 34 provinces and autonomous regions, and have a large number of customers. Currently, China Union Pay has a rapid development and new business has been introduced. The rapid growth of user scale has led to queries and complaints of related businesses also increasing significantly, which directly contributes to the pressure of Union Pay customer service department. The customer service staffs have to answer the same and simple question again and again, if only human services is applied, which directly contributes to the working pressure of the customer service staff. If the consulting business is busy, the time of customers waiting for service will increase, which will reduce customers' satisfaction with Union Pay service. In order to effectively reduce the pressure of customer service and reduce labor costs, this paper presents research on Union Pay customer service Answering System, and implements Union Pay customer service Answering System which can answer queries of users with natural language.System receives the queries raised by users. System will lookup users' questions from precreated FAQ(Frequently Asked Questions) database and judges whether the questions already exist. If users' questions exist in the FAQ database, the QA system will return the corresponding answers to the users, which is the simplest and most efficient situation. If users' questions do not exist in the FAQ database, the situation would be much more complicated. System will understand questions raised by users, and then use keywords obtained by question understand module for information retrieval with highly efficient search engine. We take the first sixty search results page as documents collection for answer extraction. The system will select answers in the documents collection which is obtained by information retrieval module. QA system will consider most similar five candidate sentences as answer and return answer to the user.In this paper, the design for the similarity of questions is not simply use ED or LCS. This paper designs the formula which meet the requirements of the similarity properties but using their respective advantages of these two methods. In order to improve the responsiveness and efficiency of the system, this paper uses Lucene retrieval system which is a powerful and cross-platform structure to create an index on the FAQ corpus. This paper uses the TF-IDF algorithm for the implement of keyword extraction. A word have larger TF-IDF value shows that the word is more important and the word has a higher weight. In this paper, the method for Chinese Word Segmentation is Maximum Matching. Taking the saving for memory space and the improvement for search efficiency into considerations, this paper use Trie tree which is a highly efficient Prefix text processing structure to store word dictionary. In this paper, Multithreading programming techniques is applied for writing web crawling program. Multithreading can significantly improve the efficiency of capturing data and reduce the system's response time to user input. Finally, the vector space model method is used to achieve the answer extraction.After many tests, system meets the basic needs of users. If the similarity threshold between user's questions and some questions in FAQ database not less than 0.5, this show users' questions can match FAQ database. In the case of match, the system's accuracy rate is greater than 85%. If the users' questions do not match any question in FAQ database, answers is extracted from the document collection. Users generally satisfy the answers extracted from document collections. If the network is not blocked, the response time of the system for user input less than one second. |
PDF Full Text Request