# One System to Rule them All: a Universal Intent Recognition System for Customer Service Chatbots

**Authors:** Juan Camilo Vasquez-Correa, Juan Carlos Guerrero-Sierra, Jose Luis Pemberty-Tamayo, Juan Esteban Jaramillo, Andres Felipe Tejada-Castro  
**arXiv:** [2112.08261](https://arxiv.org/abs/2112.08261)  
**Date:** 2021-12-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.08261v1/

## Key points

- Customer service chatbots provide information about products and services offered by companies
- Intent recognition is crucial for chatbot systems to understand and respond to customer queries
- Universal intents in customer service chatbots include salutation, switching to a human agent, farewells, and others
- A universal intent recognition system is proposed to optimize the training process of specific chatbots
- Word-embedding models like word2vec and BERT, along with deep classifiers, are used to train the model
- The model achieves a balanced accuracy of up to 80.4% in recognizing universal intents
- Misclassification errors often occur between intents with similar semantic fields
- The intent recognition system can assist in optimizing the training process of customer service chatbots and serve as a base model for more specific chatbots
- Different intents have varying request lengths, with some being short (salutation, yes/no responses) and others longer (complaints & claims)
- Models were validated using a dataset labeled with 46 different intents, showing high accuracies for most intents except under-sampled classes

## Summary

Customer service chatbots are conversational systems that provide information to customers about products and services offered by different companies. Intent recognition is a crucial component of these chatbot systems, as it allows them to understand and respond appropriately to customer queries. While there are various intents that a chatbot can be trained to recognize, there is a set of universal intents that are common across all customer service chatbots. These universal intents include salutation, switching the conversation to a human agent, farewells, and others. To optimize the training process of specific customer service chatbots, we propose the development of a universal intent recognition system. This system will be trained to recognize a selected group of 11 intents that are common in 28 different chatbots. We utilize state-of-the-art word-embedding models such as word2vec and BERT, along with deep classifiers based on convolutional and recurrent neural networks, to train the proposed model. The results show that our model is able to accurately discriminate between these universal intents with a balanced accuracy of up to 80.4%. Furthermore, the model performs equally well in recognizing intents expressed in both short and long text requests. However, misclassification errors often occur between intents with similar semantic fields, such as farewells and positive comments. The proposed intent recognition system will greatly assist in optimizing the training process of customer service chatbots by already detecting some of the available intents. Additionally, this approach can serve as a suitable base model for training more specific chatbots through transfer learning strategies. In terms of data analysis, we examined the distribution of tokens for each intent in the corpus. Some intents have very short requests like salutation, yes/no responses, farewells, and FAQs while petitions, complaints & claims (PCC), negative comments and requests for attention lines tend to have longer requests due to customers expressing dissatisfaction or describing bad experiences with the service. We also validated our models using the Natural Language Data for Human-Robot Interaction in Home Domain dataset which consists of 25 716 requests labeled with 46 different intents; showing high accuracies for most intents except for under-sampled classes like help and negative comments where farewells were often misclassified as positive comments due to their similar semantic fields related to acknowledging provided services at end conversations .