tuanbc88 / open-intent-discovery

Unsupervised clustering for open intent discovery

Open Intent Discovery through Unsupervised Semantic Clustering and Dependency Parsing

Introduction

Intent understanding plays an important role in dialog systems, and is typically formulated as a supervised learning problem. However, it is challenging and time-consuming to design the intents for a new domain from scratch, which usually requires a lot of manual effort of domain experts. This project presents an unsupervised two-stage approach to discover intents and generate meaningful intent labels automatically from a collection of unlabeled utterances in a domain, as illustrated in the following figure.

In the first stage, we aim to generate a set of semantically coherent clusters where the utterances within each cluster convey the same intent. We obtain the utterance representation from various pre-trained sentence embeddings and apply clustering methods. In the second stage, the objective is to generate an intent label automatically for each cluster. We extract the ACTION-OBJECT pair from each utterance using a dependency parser and take the most frequent pair within each cluster, e.g., book-restaurant, as the generated intent label. We empirically show that the proposed unsupervised approach can generate meaningful intent labels automatically and achieve high precision and recall in utterance clustering and intent discovery.

Source Code

This repository contains the core code for running the experiments. The SNIPS dataset is preprocessed from https://github.com/sonos/nlu-benchmark. Please cite their paper if you use the dataset.

How to run the experiments?

The batch script can be run as follows:

bash batch.sh

Citation

If you use the released source code in your work, please cite the following paper:

@article{liu2021open,
  title={Open Intent Discovery through Unsupervised Semantic Clustering and Dependency Parsing},
  author={Liu, Pengfei and Ning, Youzhang and Wu, King Keung and Li, Kun and Meng, Helen},
  journal={arXiv preprint arXiv:2104.12114},
  year={2021}
}

Report

Please feel free to create an issue or send emails to the first author at [email protected].