【作者】 黄震；

【导师】 吴含前； 宛斌； 彭鹏；

【作者基本信息】 东南大学 ， 软件工程（专业学位）， 2021， 硕士

【摘要】 近年来,伴随着互联网金融冲击,金融行业竞争日益激烈,传统的银行业面临着较大的挑战。在银行智能化改革的大趋势下,如何应用人工智能技术在提高效率的同时提升客户体验是银行需要面对和解决的难题,其中客服问答是核心难点之一。本文构建了银行客服和理财相关的问答数据库,并由此实现了系统核心的问答功能:对于用户输入的问题,根据短文本匹配方法来直接返回最合适的答案。同时系统也实现了理财相关的扩展服务功能,在提升客户体验的同时,满足了客户日益增长的投资理财需求。面向银行客服的智能问答系统主要由后台管理模块、智能问答模块和扩展服务模块组成,其中智能问答模块是系统的核心模块。系统的问答功能主要由候选问题集快速召回和短文本匹配精确排序两个过程实现,召回过程使用无监督的文本向量表示,排序过程使用有监督的基于BERT的孪生网络模型。后台管理模块用于问答数据集的维护更新和用户评价数据的分析。扩展服务模块主要实现了理财相关的功能,具体包括个人风险评估和个人理财推荐。本文设计的智能问答系统框架可以看作一个基于文本向量表示的通用的问答系统框架,对于其他领域的智能问答系统应用具有较大参考价值。对问答系统中的短文本匹配进行测试,结果表明系统设计的召回+排序的解决方案有效的提高了短文本匹配的效率,问答时延小于1秒,问答准确率大于90%。系统的功能和非功能测试表明,该系统实现了高效的问答功能和后台管理功能,在银行客服领域中具有较大的应用价值。更多还原

【Abstract】 Over the past years,the competition in the financial industry has become increasingly fierce,and the traditional banking industry is facing greater challenges with the impact of Internet finance.Under the trend of intelligent banking reform,it is a problem that how to apply artificial intelligence technology to improve efficiency and customer experience,which banks need to face and solve,and Q&A of customer service is one of the core difficulties.A Q&A database related to bank customer service and financial management is constructed in this study,and it implements the core function of the system that to directly return the most appropriate answer for user’s question according to the short text matching method.Meanwhile,the system also implements the extended service function related to financial management,which not only improves the customer experience,but also meets the growing investment and financial management needs of customers.The intelligent question answering system for bank customer service is mainly composed of background management module,intelligent question answering module and extended service module.The intelligent question answering module is the core module of the system.The question answering function of the system is mainly composed by two steps that quick recall of candidate question set and accurate sorting of short text matching.The recall process uses unsupervised text vector representation,and the sorting process uses a supervised twin network model based on BERT.The background management module is designed for the maintenance and update of Q&A database and the analysis of user’s evaluation data.The extended service module mainly implements the functions related to financial management,including personal risk assessment and personal financial recommendation.The intelligent question answering system framework designed in this paper can be regarded as a general question answering system framework based on text vector representation,which has a great reference value for the application of intelligent question answering system in other fields.The test result of the short text matching in question answering system shows that the solution of recall and sorting effectively improves the efficiency of short text matching,the question answering delay is less than one second,and the accuracy of the question answering module is over 90%.The functional and non-functional test of the system shows that the system implements question answering function and background management function,which is efficient and has more significance in bank customer service.更多还原