一种面向移动终端的自然口语任务理解方法① - 计算机系统应用.pdfVIP

计 算 机 系 统 应 用 2013 年 第 22 卷 第 8 期 ① 一种面向移动终端的自然口语任务理解方法 1 2 1 2 郭 群 , 李剑锋 , 陈小平 , 胡国平 1( 中国科学与技术大学 计算机科学与技术学院, 合肥 230027) 2(安徽科大 讯飞信息科技股份有限公司研究院, 合肥 230088) 摘 要: 随着移动互联时代的到来和语音识别技术的日益成熟, 通过语音的交互方式来使用移动终端成为一种 趋势. 如何理解用户自然状态下的口语输入, 传统的做法是手写上下文无关的文法规则, 但是文法规则的书写需 耗费大量的人力和物力, 很难去维护和更新. 提出一种采用支持向量机和条件随机场串行结合的方法, 把口语任 务理解分解为任务发现和信息抽取两个过程, 并最终将任务表达成语义向量的形式. 最终对“讯飞语点”语音助手 用户返回的八个不同的任务种类的数据进行了测试, 在一比一的噪声中识别任务语义表达的准确率为 90.29%, 召回率为 88.87%. 关键词: 口语理解; 任务发现; 信息抽取; 支持向量机; 条件随机场 A Method to Understand Spontaneous Spoken Tasks for Mobile Terminals 1 2 1 2 GUO Qun , LI Jian-Feng , CHEN Xiao-Ping , HU Guo-Ping 1(School of Electronic Science and Technology, University of Science and Technology of China, Hefei 230027, China) 2(Research Center, Anhui USTC iFLYTEK Co. Ltd, Hefei 230088, China) Abstract: With the development of mobile Internet and automatic speech recognition (ASR), the mobile terminal through voice interaction has become a trend. The traditional method to understand users spontaneous spoken language is to write context-free grammars(CFGs)manually. But it is laborious and expensive to construct a grammar with good coverage and optimized performance, and difficult to maintain and update. We proposed a new approach to spoken language understanding combining support vector machine(SVM)and conditional random fields(CRFs), which detect task and extract task semantic information from spontaneous speech input respectively. Tasks are represented as a vector of task name and semantic information. Eight different tasks from“iFLYTEK yudian”voice mobile assistant are