Context-sensitive and role-dependent spoken language understanding using bidirectional and attention LSTMs

To understand speaker intentions accurately in a dialog, it is important to consider the context of the surrounding sequence of dialog turns. Furthermore, each speaker may play a different role in the conversation, such as agent versus client, and thus features related to these roles may be important to the context. In previous work, we proposed context-sensitive spoken language understanding (SLU) using role-dependent long short-term memory (LSTM) recurrent neural networks (RNNs), and showed improved performance at predicting concept tags representing the intentions of agent and client in a human-human hotel reservation task. In the present study, we use bidirectional and attention-based LSTMs to train a roledependent context-sensitive model to jointly represent both the local word-level context within each utterance, and the left and right context within the dialog. The different roles of client and agent are modeled by switching between role-dependent layers. We evaluated label accuracies in the hotel reservation task using a variety of models, including logistic regression, RNNs, LSTMs, and the proposed bidirectional and attentionbased LSTMs. The bidirectional and attention-based LSTMs yield significantly better performance in this task.