Intent Discovery With Or Without Labeled Data Using Dependency Parser

Anonymous

Keywords: Intent Discovery, KMeans Clustering

TL;DR: A clustering and evaluation approach that can be used in semi-supervised or unsupervised modes for new intent discovery.

Abstract:

In dialogue applications, machine learning classification models are often used to classify user utterances into different intents that help to understand the users.
In real world scenarios, however, some utterances may not belong to any of the anticipated intent categories.
Furthermore, supervised classification models are not a viable solution when data of a new domain is introduced without the corresponding labels.
In this work, we present a clustering and evaluation approach that can be used in semi-supervised or unsupervised modes, depending on the (non-)availability of training data for new intent discovery.
This method assigns meaningful intent-labels by determining the optimal number of clusters and evaluating the performance of the clustering results. In addition, it assigns a TF-IDF score to individual samples within a cluster.