【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。更多还原

【Abstract】 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.更多还原