- Abstract: Human language is distinguished by powerful semantics, rich structure, and incredible flexibility. It enables us to communicate with each other, thereby affecting the decisions we make and actions we take. While Artificial Intelligence (AI) has made great advances both in sequential decision-making using Markov Decision Processes (MDPs) and in Natural Language Processing (NLP), the potential of language to inform sequential decision-making is still unrealized. We explore how the different functional elements of natural language---such as verbs, nouns and adjectives---relate to decision process formalisms of varying complexity and structure. We attempt to determine which elements of language can be usefully grounded to a particular class of decision process and how partial observability changes the usability of language information. Our work show that more complex, structured models can capture linguistic concepts that simple MDPs cannot. We argue that the rich structure of natural language indicates that reinforcement learning should focus on richer, more highly structured models of decision-making.
- Keywords: language, reinforcement learning