对话意图及语音识别错误对交互体验的影响

在自然人机对话中，由于环境噪声、方言口音等因素带来的语音识别错误以及语义分析的不充分等原因，计算机在理解用户交互意图时出现偏差，使得计算机对要反馈的话题出现错误，造成人机对话进程的断裂.以面向咖啡为主题的漫谈式人机对话为例，将对话中断分为3种情况：话题反馈不当引起中断、话题正确情况下的模糊反馈不当和精确反馈不当引起中断.根据用户与计算机对话的记录分析比较上述3种情况下人机对话进程断裂情况.统计数据结果表明，话题反馈不当带来的对话中断最为明显，在对话进程断裂情况中达到了60.1%的比例；在话题反馈正确情况下，模糊回答不当和精确回答不当带来的话题中断比例分别为22.2%和21.6%；在语音识别错误情况下，语义分析会带来数量更大的反馈错误.实验数据分析结果表明，在语音识别错误情况下，根据上下文信息提高计算机对用户话题反馈的准确率，能够有效降低人机对话的中断，提高人机对话的自然度.该工作为自然人机对话的意图分类重要性提供了数据分析和实验论证.

Error Analysis of Intention Classification and Speech Recognition in Human-Computer Dialog

In the natural human-computer dialogue system, environmental noises, accents and some other factors may cause the speech recognition errors which leads to computers'' error responses to human. The dialogs are often interrupted by the system''s bad responses. Three types of human computer interruptions are considered in this paper:improper feedback for topic, improper response for a vague user query, and improper feedback for an exact user query. According to the records of the user and computer dialogue analysis, the interruptions caused by three situations above are compared and used to analyze the importance of intention classification in human-computer conversation. The statistical data find that the dialogue interruption caused by the inappropriate topic feedback is the most obvious problem, amounting to 60.1%. Under the correct feedback of the topic, the interrupt ratio of the subject caused by accurate answer and fuzzy answer is 22.2% and 21.6% respectively. In the case of error speech recognition, semantic analysis can bring more feedback error to the error of speech recognition. The analysis of experimental data shows that the speech recognition errors, can effectively reduce the man-machine conversation interrupt and improve the naturalness of human-computer dialogue system according to the context information to improve the accuracy of the computer on the topic of user feedback,. This paper provides the importance of intention classification in human machine dialogue, which helps to improve the performance of human-computer dialogue system.