Reinforcement Learning for Improving Coherence of Multi-turn Responses in Deep Learning-Based Chatbots

Abstract

Chatbots are still far behind in their ability to hold meaningful conversations. The objective of the work is to implement and improve the multi-turn responses of deep learning-based chatbots. Multi-turn response is the ability of a chatbot to give coherent and sensible responses in successive turns. Firstly, sequence to sequence (Seq2Seq) model was built, and its responses were analyzed by varying training parameters. Secondly, the reinforcement learning (RL) method using the Seq2Seq model was implemented, and it is demonstrated that this improves coherence in multi-turn conversations. The RL model performed better than the Seq2Seq model in terms of BiLingual Evaluation Understudy (BLEU) score with a score of 0.3334 compared to 0.2336 of the Seq2Seq model. Average conversation length was found to increase with RL with 3.75 turns compared to 3.05 turns with Seq2Seq.