A Deep Reinforcement Learning Chatbot

TLDR

The system’s machine‑learning architecture suggests it will improve as more data are collected. The authors present MILABOT, a deep reinforcement learning chatbot developed by MILA for the Amazon Alexa Prize competition. The system is an ensemble of natural‑language generation and retrieval models—template‑based, bag‑of‑words, sequence‑to‑sequence, and latent‑variable neural networks—trained via reinforcement learning on crowdsourced and real‑world user data to select appropriate responses. MILABOT can converse on popular small‑talk topics via speech and text, and A/B testing with real‑world users showed it performed significantly better than many competing systems.

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.