Encoding Context in Task-Oriented Dialogue Systems Using Intent, Dialogue Acts, and Slots

Abstract

Extracting context from natural language conversations has been the focus of applications which communicate with humans. Understanding the meaning and the intent of the user input, and formulating responses based on a contextual analysis mimicking that of an actual person is at the heart of modern-day chatbots and conversational agents. For this purpose, dialogue systems often use context from previous dialogue history. Thus, present-day dialogue systems typically parse over user utterances and sort them into semantic frames. In this paper, a bidirectional RNN with LSTM and a CRF layer on top is used to classify each utterance into its resultant dialogue act. Furthermore, there is a separate bidirectional RNN with LSTM and attention for the purpose of slot tagging. Slot annotations use the inside-outside-beginning (IOB) scheme. Softmax regression is used to determine the intent of the entire conversation. The approach is demonstrated on data from three different domains.