基于逻辑回归的口语理解方法研究  

Research on logistic regression spoken language understanding method

作　　者：张蕤[1] 孙甲松[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《信息技术》2016年第4期92-95,104,共5页Information Technology

摘　　要：针对语音识别错误导致口语理解系统性能下降的问题,提出一种易于训练且解码快速的鉴别式口语理解方法。首先为每个语义要素建立一个二类逻辑回归模型,随后根据领域中的限制关系建立联合概率模型。在英语公开数据集DSTC2上的实验结果表明,该方法优于人工规则方法和语义元组分类器模型。Spoken language understanding（ SLU） suffers from system＇s performance degradation which caused by automatic speech recognition＇s errors. This paper presents a discriminative method for SLU which is easier on training and faster on decoding. The method first trains a binary logistic regression model for each semantic item. After that, a joint probabilistic model with domain restrictions is established. The experimental results on DSTC2 corpus show that it outperforms hand-crafted parser method and semantic tuple classifier model.

关 键 词：口语对话系统 口语理解 逻辑回归 语义框架 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]