Interactive language learning - How to explore complex environments using natural language?

Implicit knowledge about the physical world we live in is gained almosteffortlessly through interaction with the environment. In the same manner, this knowledge cannot be simply inferred from language, as humans normally avoid stating what is trivially implied or observed in the world. This paper is about a novel perspective into progressing artificial intelligence toward understanding the true language meaning through interaction with complex environments. The arising field of text-based games seems to hold the key for such an endeavour. Text-based games placed in a reinforcement learning formalism have the potential of being a strategic path into advancing real-world natural language applications - the human world itself is one of partial understanding through communication and acting on the world using language. We present a comparative study highlighting the importance of having a unified approach in the area of learning agents to play families of text-based games,with the scope of establishing a benchmark that will enable the community to advance the state of the art. To this end, we will look at the corpora and the first two winner solutions from the competition launched by Microsoft Research - FirstTextWorld Problems. The games from the proposed corpora share the same objective, cooking a meal after collecting ingredients from a modern house environment, having the layout and the recipes change from one game to another.