Finding Dominant User Utterances And System Responses in Conversations

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 723–732, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Finding Dominant User Utterances And System Responses in Conversations Dhiraj Madan and Sachindra Joshi IBM Research Labs New Delhi, India {dhimadan,jsachind@in.ibm.com} Abstract There are several dialog frameworks which allow manual specification of in- tents and rule based dialog flow. The rule based framework provides good control to dialog designers at the expense of be- ing more time consuming and laborious. The job of a dialog designer can be re- duced if we could identify pairs of user intents and corresponding responses au- tomatically from prior conversations be- tween users and agents. In this paper we propose an approach to find these frequent user utterances (which serve as examples for intents) and corresponding agent re- sponses. We propose a novel SimCluster algorithm that extends standard K-means algorithm to simultaneously cluster user utterances and agent utterances by tak- ing their adjacency information into ac- count. The method also aligns these clus- ters to provide pairs of intents and re- sponse groups. We compare our results with those produced by using simple K- means clustering on a real dataset and ob- serve upto 10% absolute improvement in F1-scores. Through our experiments on synthetic dataset, we show that our algo- rithm gains more advantage over K-means algorithm when the data has large vari- ance. 1 Introduction There are several existing works that focus on modelling conversation using prior human to hu- man conversational data (Gaˇsi´c et al., 2013; Young et al., 2013; Henderson et al., 2014). (Higashinaka et al., 2011) models the conversation from pairs of consecutive tweets. Deep learning based ap- proaches have also been used to model the di- alog in an end to end manner (Vinyals and Le, 2015; Serban et al., 2015). Memory networks have been used by Bordes et al (2016) to model goal based dialog conversations. More recently, deep reinforcement learning models have been used for generating interactive and coherent dialogs (Li et al., 2016) and negotiation dialogs (Lewis et al., 2017). Industry on the other hand has focused on build- ing frameworks that allow manual specification of dialog models such as api.ai1, Watson Conversa- tional Services2, and Microsoft Bot framework3. These frameworks provide ways to specify intents, and a dialog flow. The user utterances are mapped to intents that are passed to a dialog flow manager. The dialog manager generates a response and up- dates the dialog state. See Figure 1 for an exam- ple of some intents and a dialog flow in a techni- cal support domain. The dialog flow shows that when a user expresses an intent of # laptop heat, then the system should respond with an utterance “Could you let me know the serial number of your machine ”. The designer needs to specify intents (for example # laptop heat, # email not opening) and also provide corresponding system responses in the dialog flow. This way of specifying a …

Content truncated to 500 words