Abstract: The semantic frame based spoken language understanding involves two decisions—key semantic concept recognition and intent determination. This paper focuses on execution strategies of two decisions. Firstly, this paper studies independent and cascade strategies. Then, it proposes joint model, which is based on triangular-chain conditional random fields to build a joint model between key semantic concept recognition and intent determination to denote their dependence relationship. Experiments about these strategies show that the performance of joint model for key semantic concept recognition is better than other strategies.

Key words: spoken language understanding, triangular-chain conditional random field, key semantic concept recognition, intent determination, execution strategies

摘要： 口语理解的语义框架包括两个决策——关键语义概念识别和意图识别，主要针对这两个决策的执行策略进行研究。首先研究了并联型和级联型两种策略；然后在此基础上提出了联合型结构进行中文口语理解，即通过三角链条件随机场对意图以及关键语义概念共同建模，用一个单独的图模型结构共同表示它们的依赖关系。通过与其他几种策略进行比较实验得出结论：该模型可以将两个任务一次完成，在关键语义概念识别任务上性能优于其他的执行策略。

关键词: 口语理解, 三角链条件随机场, 关键语义概念识别, 意图识别, 执行策略