Finding Dominant User Utterances And System Responses in Conversations

• DOI: 10.48550/arxiv.1710.10609
• Publication Date: 2017-01-01
• Authors: Dhiraj Madan; Sachindra Joshi
• Keywords: Dialog system; Adjacency list; K-Means Clustering

There are several dialog frameworks which allow manual specification of intents and rule based flow. The framework provides good control to designers at the expense being more time consuming laborious. job a designer can be reduced if we could identify pairs user corresponding responses automatically from prior conversations between users agents. In this paper propose an approach find these frequent utterances (which serve as examples for intents) agent responses. We novel SimCluster algorithm that extends standard K-means simultaneously cluster by taking their adjacency information into account. method also aligns clusters provide response groups. compare our results with those produced using simple Kmeans clustering on real dataset observe upto 10% absolute improvement in F1-scores. Through experiments synthetic dataset, show gains advantage over when data has large variance.