欢迎访问中国科学院大学学报,今天是

中国科学院大学学报 ›› 2015, Vol. 32 ›› Issue (2): 252-258.DOI: 10.7523/j.issn.2095-6134.2015.02.015

• 信息与电子科学 • 上一篇    下一篇

航班预定口语对话系统的设计与实现

陈振锋1,2, 杨晓昊3, 吴蔚澜1,2, 刘加3, 夏善红2   

  1. 1. 中国科学院电子学研究所 传感技术国家重点实验室, 北京 100190;
    2. 中国科学院大学, 北京 100190;
    3. 清华大学电子工程系 清华信息科学与技术国家实验室, 北京 100084
  • 收稿日期:2014-03-24 修回日期:2014-05-16 发布日期:2015-03-15
  • 通讯作者: 夏善红, shxia@mail.ie.ac.cn
  • 基金资助:

    国家自然科学基金(61005019,61273268,90920302)和北京市自然科学基金(KZ201110005005)资助

Design and implementation of mandarin spoken dialogue system for flight reservation

CHEN Zhenfeng1,2, YANG Xiaohao3, WU Weilan1,2, LIU Jia3, XIA Shanhong2   

  1. 1. State Key Laboratory on Transducing Technology, Institute of Electronics, Chinese Academy of Sciences, Beijing 100190, China;
    2. University of Chinese Academy of Sciences, Beijing 100190, China;
    3. Tsinghua National Laboratory for Information Science and Technology, Department of Electronic Engineering, Tsinghua University, Beijing 100084, China
  • Received:2014-03-24 Revised:2014-05-16 Published:2015-03-15

摘要:

介绍一个航班预定口语对话系统的设计与实现, 该系统允许用户通过普通话进行航班信息查询与预定.重点介绍口语对话系统中的口语语言理解.为了克服语音识别引入的识别错误导致语义理解错误的问题, 提出基于词混淆网络的两阶段中文口语语言理解方法:首先从词混淆网络中选择N元文法作为分类特征, 进行主题分类, 并通过语义分类模型解析获取对应的语义树结构;然后利用基于规则的语义槽填充器抽取相应的语义槽属性-值.该方法是数据驱动的, 训练数据的标记比较容易.实验在汉语航班预定领域进行, 结果表明, 在语音识别字错误率很高的情况下, 该方法比传统的基于语法规则的语言理解方法更加鲁棒, 在语义理解正确率方面有明显改善.

关键词: 口语对话系统, 口语语言理解, 语义理解, 词混淆网络, 对话管理

Abstract:

We present a spoken dialogue system for flight reservation, which allows users to inquire information about flight in mandarin. We describe the design and the implementation of our system, focusing on spoken language understanding (SLU). Considering that the speech recognizer inevitably makes errors, we propose a new two-stage mandarin SLU approach based on word confusion network. Firstly, the semantic tuple classifier is used to identify the topic of an input utterance using N-gram features extracted from the word confusion network and to parse a semantic tree by recursively calling semantic classification models. Then the rule-based semantic slot filler is used to extract the corresponding slot/value pairs. The advantage of the proposed approach is that it is mainly data-driven and requires minimally annotated corpus for training. Experiment has been carried out in the Chinese flight reservation domain, which shows that the proposed approach is more robust to speech recognition errors than the conventional handcrafted rule-based parser, and substantially improves performance of accuracy when the ASR word error rate is high.

Key words: spoken dialogue system, spoken language understanding, semantic analysis, word confusion network, dialogue management

中图分类号: