口语对话系统中的一种稳健语言理解算法

Robust language understanding algorithm in spoken dialogue systems

为提高口语对话系统中语言理解的稳健性,提出了一种基于最大后验统计框架的两级搜索的理解算法。第一级用概念捆绑达到提取句中关键成分并剔除某些干扰成分的目的;第二级采用改进的基于树扩展的稳健句法分析搜索最佳理解结果,同时引入用户意图推断和句子特征短语两方面的信息对搜索空间进行约束,进一步提高了理解的稳健性和实时率。实验表明,该算法应用于火车信息查询领域,在0.22倍实时下,能得到13.6%的句意理解错误率和25.4%的概念理解错误率,相对基线系统分别为降低了23.2%和9.3%。

A two-level understanding algorithm was developed to improve the robustness of language understanding in spoken dialogue systems. In the first level, arcs on a word graph are combined to form a concept graph, which eliminates some error words from the process. In the second level, a graph expansion-based robust parser is used to search the concept graph for the best understanding result within a statistical framework. To further improve the robustness and efficiency, user plan inferences and key phrases are introduced to constrain the search space. Experimental results show that the algorithm achieves a sentence meaning error rate of 13.6% and a concept error rate of 25.4% with a 0.22 real time factor. The relative error rate reductions are 23.2% and 9.3%, respectively.