Abstract: Most of the previous studies explored Spoken Language Understanding（SLU） framework for building separate model for each task, such as slot filling or intent detection, ignoring the mutual promotion of this tasks. A BLSTM-CNN-CRF architecture for joint modeling of slot filling and intent detection is proposed based on the advantages of deep learning. Each word-level label is greedily determined by the hidden representation from the Bi-directional Long Short-Term Memory（BLSTM） network in each step. The semantic features of the whole utterance are automatically extracted through Convolutional Neural Network（CNN）. Finally, the output vectors of them are fed to the Conditional Random Fields（CRF） layer to jointly decode the best label sequence. The joint model achieves competitive performance on the benchmark Airline Travel Information System（ATIS） task without any artificial features.

Key words: slot filling, intent detection, joint model, deep learning

摘要： 口语语言理解（SLU）中的槽填充和意图识别任务通常是分别进行建模，忽略了任务之间的关联性。基于深度学习优势提出一种BLSTM-CNN-CRF学习框架，为槽填充和意图识别任务构建联合模型。双向长短期记忆网络（BLSTM）对全句的单词标签进行标注，卷积神经网络（CNN）用以提取全句的语义特征，条件随机场（CRF）通过解码单词标签与语义特征，获得全句的最佳序列标签。在航空旅行信息系统（ATIS）数据集上的实验表明，联合模型在不依赖于任何人工特征的情况下获得较高性能。

关键词: 槽填充, 意图识别, 联合模型, 深度学习