A Deep Reinforcement Learning Chatbot

TL;DR

This paper introduces MILABOT, a deep reinforcement learning chatbot designed for engaging small talk, combining multiple natural language models and trained via reinforcement learning to improve responses based on user interactions.

Contribution

The paper presents a novel ensemble-based deep reinforcement learning approach for chatbot response selection, optimized through real-world user feedback and crowdsourced data.

Findings

MILABOT outperforms many competing systems in A/B tests.

The system improves with additional data due to its machine learning architecture.

Ensemble of diverse models enhances conversational capabilities.

Abstract

We present MILABOT: a deep reinforcement learning chatbot developed by the Montreal Institute for Learning Algorithms (MILA) for the Amazon Alexa Prize competition. MILABOT is capable of conversing with humans on popular small talk topics through both speech and text. The system consists of an ensemble of natural language generation and retrieval models, including template-based models, bag-of-words models, sequence-to-sequence neural network and latent variable neural network models. By applying reinforcement learning to crowdsourced data and real-world user interactions, the system has been trained to select an appropriate response from the models in its ensemble. The system has been evaluated through A/B testing with real-world users, where it performed significantly better than many competing systems. Due to its machine learning architecture, the system is likely to improve with additional data.