Dialog Intent Induction with Deep Multi-View Clustering

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4016–4025, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4016 Dialog Intent Induction with Deep Multi-View Clustering Hugh Perkins and Yi Yang ASAPP Inc. New York, NY 10007 {hp+yyang}@asapp.com Abstract We introduce the dialog intent induction task and present a novel deep multi-view cluster- ing approach to tackle the problem. Dialog intent induction aims at discovering user in- tents from user query utterances in human- human conversations such as dialogs between customer support agents and customers.1 Mo- tivated by the intuition that a dialog intent is not only expressed in the user query utter- ance but also captured in the rest of the dialog, we split a conversation into two independent views and exploit multi-view clustering tech- niques for inducing the dialog intent. In par- ticular, we propose alternating-view k-means (AV-KMEANS) for joint multi-view represen- tation learning and clustering analysis. The key innovation is that the instance-view rep- resentations are updated iteratively by predict- ing the cluster assignment obtained from the alternative view, so that the multi-view rep- resentations of the instances lead to similar cluster assignments. Experiments on two pub- lic datasets show that AV-KMEANS can in- duce better dialog intent clusters than state- of-the-art unsupervised representation learn- ing methods and standard multi-view cluster- ing approaches.2 1 Introduction Goal-oriented dialog systems assist users to ac- complish well-defined tasks with clear intents within a limited number of dialog turns. They have been adopted in a wide range of applications, in- cluding booking flights and restaurants (Hemphill et al., 1990; Williams, 2012), providing tourist in- formation (Kim et al., 2016), aiding in the cus- tomer support domain, and powering intelligent 1We focus on inducing abstract intents like BookFlight and ignore detailed arguments such as departure date and destination. 2When ready, the data and code will be pub- lished at https://github.com/asappresearch/ dialog-intent-induction. Customer 1: A wireless charging case is fancy and all but can we get a “find my airpod” feature going? Agent 1: If you have lost your AirPods, Find My iPhone can help you locate them. Customer 2: hey man I lost and miss my airpods plz help me! Agent 2: Hi there! With iOS 10.3 or later, Find My iPhone can help you locate missing AirPods. Figure 1: Two dialogs with the FindAirPods user intent. The user query utterances of the two dialogs are lexically and syntactically dissimilar, while the rests of the dialogs are similar. virtual assistants such as Apple Siri, Amazon Alexa, or Google Assistant. The first step to- wards building such systems is to determine the target tasks and construct corresponding ontolo- gies to define the constrained set of dialog states and actions (Henderson et al., 2014b; Mrkˇsi´c et al., 2015). Existing work assumes the target tasks are given and excludes dialog intent discovery from the dia- log system design pipeline. Because of this, most of the works focus on …

Content truncated to 500 words