Adaptive natural language generation in dialogue using reinforcement learning

This paper presents a new model for adaptive Natural Language Generation (NLG) in dialogue, showing how NLG problems can be approached as statistical planning problems using Reinforcement Learning. This approach brings a number of theoretical and practical benefits such as finegrained adaptation, generalization, and automatic (global) optimization. We present the model and related work in statistical/trainable NLG, discuss its applications, and provide a demonstration of the approach, showing policy learning for adaptive information presentation decisions (Contrast, Cluster, or List items). An adaptive NLG policy learned in our framework shows a statistically significant 27% relative increase in reward over an “RL-majority” baseline policy for the same task. We thereby also show that that such NLG problems should be approached in combination with dialogue management decisions, and we show how to jointly optimize NLG and dialogue management plans.