# Intent Mining from past conversations for Conversational Agent

**Authors:** Ajay Chatterjee, Shubhashis Sengupta  
**arXiv:** [2005.11014](https://arxiv.org/abs/2005.11014)  
**Date:** 2020-05-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2005.11014v1/

## Key points

- Framework for intent mining from past conversations to improve conversational systems
- Chatbots used for round-the-clock support and customer engagement
- Challenges in building and training intent models due to lack of diverse training data
- Four-step intent discovery framework proposed by the authors
- Extract textual utterances using a pre-trained Dialog Act Classifier
- Automatically cluster similar user utterances using ITER-DBSCAN algorithm
- Manual annotation of clusters with intent labels by subject matter experts
- Propagation of intent labels to unmapped utterances
- Introduction of better sentence representation method using Universal Sentence Encoder
- Effectiveness demonstrated using Microsoft application IT support data and publicly available datasets
- Comprehensive framework improves coverage, accuracy, and saves time compared to manual annotation methods
- Potential applications beyond conversational systems in short text clustering and labeling tasks
- Future work includes scalability for larger datasets and exploration in general-purpose clustering tasks

## Summary

In this paper, the authors present a framework for intent mining from past conversations to improve conversational systems. They highlight that chatbots are increasingly being used to provide round-the-clock support and enhance customer engagement. However, building and training an intent model for these systems can be challenging due to the need for substantial and diverse training data. To address these issues, the authors propose a four-step intent discovery framework. The first step involves extracting textual utterances from conversations using a pre-trained domain agnostic Dialog Act Classifier. Then, similar user utterances are automatically clustered using a novel density-based clustering algorithm called ITER-DBSCAN. Next, subject matter experts manually annotate the clusters with intent labels. Finally, intent labels are propagated to the utterances that are not mapped to any cluster. The authors also introduce a better sentence representation method using Universal Sentence Encoder for identifying imbalanced classes. They demonstrate the effectiveness of their framework using internal Microsoft application IT support conversational data and publicly available datasets for intent classification and short text classification. Overall, this paper presents a comprehensive framework for intent mining from past conversations in conversational systems which shows promise in improving coverage of intents, accuracy, and saving time compared to manual annotation methods. Additionally, the framework has potential applications beyond conversational systems in short text clustering and labeling tasks. In terms of future work, the authors aim to make their algorithm more scalable for larger datasets as well as explore its application in general-purpose clustering tasks such as knowledge graph generation and other areas of data mining.