面向任务型多轮对话的粗粒度意图识别方法  

Coarse-grained Intent Recognition Method for Task-oriented Multi-turn Dialogue

作　　者：叶铱雷 曹斌[1] 范菁[1] 王俊[2] 陈江斌 YE Yi-lei;CAO Bin;FAN Jing;WANG Jun;CHEN Jiang-bin(College of Computer Science&Technology,Zhejiang University of Technology,Hangzhou 310023,China;China Telecom Corporation Limited Zhejiang Branch,Hangzhou 310040,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]中国电信股份有限公司浙江分公司,杭州310040

出　　处：《小型微型计算机系统》2020年第8期1620-1626,共7页Journal of Chinese Computer Systems

基　　金：国家重点研发计划项目(2018YFB1402800)资助;浙江省自然科学基金项目(LY19F020030)资助;浙江省大学生科技创新活动计划(新苗人才计划)项目(2019R403092)资助。

摘　　要：为提高客服服务水平,越来越多企业开始建立智慧客服系统.其中,用户意图识别是智慧客服系统的重要组成部分.通过监听用户和客服之间的对话,分析用户来电意图,为智慧客服系统提供执行依据.现有的意图识别方法更倾向于识别多轮对话中每句话的用户意图.在呼叫中心客服对话复杂场景下,用户会在多轮对话中表述很多与业务无关的对话,而我们并不需要关心这些对话.此外,用户单个意图可能蕴含在多轮对话内容中,仅依靠用户的单次表述无法准确识别用户意图.换句话说,现有意图识别方法的意图识别粒度过小,不适用于呼叫中心客服对话复杂场景.因此,本文提出了一种面向任务型多轮对话的粗粒度意图识别方法,将意图识别任务拆分为了对话序列标注任务和意图分类任务,解决了现有方法的不足.本文采用真实的呼叫中心电话客服对话文本集合进行实验.实验结果表明我们的方法对比于一些现有的方法具有更好的准确率.In order to improve the level of customer service,more and more companies have begun to establish an intelligent customer service system.And user intent recognition is an important part of the intelligent customer service system.By monitoring the dialogue between the user and the customer service,the user intent is analyzed to provide an execution basis for the intelligent customer service system.Existing intent recognition methods are more likely to recognize user’s intent for each sentence in a multi-turn conversation.In the complex scene of call center customer service dialogue,users will express many conversations that are not related to the business in multiple turns of dialogue.And we don’t need to care about these conversations.And the user’s single intent may be contained in multiple turns of conversation content,and the user’s intent cannot be accurately recognized by relying on the user’s single expression.In other words,the intent recognition granularity of the existing intent recognition methods is too small.And the methods are not suitable for the complex scene of the call center customer service dialogue.Therefore,we propose a coarse-grained intent recognition method for task-oriented multi-turn dialogue,which splits the intent recognition task into dialogue sequence tagging task and dialogue classification task.It solves the shortcomings of the existing methods.We experimented with a real set of call center telephone customer service conversation texts.The experimental results show that our method has better accuracy than some existing method.

关 键 词：多轮对话 粗粒度 意图识别 对话序列标注 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]