Learning Rewards From Linguistic Feedback

We explore unconstrained natural language feedback as a learning signal for artificial agents. Humans use rich and varied to teach, yet most prior work on interactive from assumes particular form of input (e.g., commands). propose general framework which does not make this assumption, instead using aspect-based sentiment analysis decompose into over the features Markov decision process. then infer teacher's reward function by regressing features, an analogue inverse reinforcement learning. To evaluate our approach, we first collect corpus teaching behavior in cooperative task where both teacher learner are human. implement three learners: sentiment-based "literal" "pragmatic" models, inference network trained end-to-end predict rewards. re-run initial experiment, pairing human teachers with these learners. All models successfully learn feedback. The approaches performance model, while model nears performance. Our provides insight information structure naturalistic linguistic well methods leverage it