ORIGINAL

Abstract

In recent years, there has been an increasing interest in text-based games in Natural Language Processing research, in specific in the area of language understanding, which aims to develop algorithms and paradigms of how a computer can develop an understanding of language on the higher level of semantics and meaning. Text-based games have been researched as a tool to achieve this goal. In these games, the agent has to take in instructions and information in the form of text and make decisions about their actions based on that. Through approaches in reinforcement learning, there have been many algorithms that successfully learn policies for such games that generalize well.

This paper will investigate the difference between understanding and memorization in neural network approaches. I utilize the TextWorld platform - a game generation engine developed by Microsoft Research that can be used to create text games of various characteristics. I will investigate the agent's ability to correctly recognize the same situation given modified information. For the purposes of this, I will work on a stochastic version of a text-based game, where at each turn the agent is given a randomized modification of the original text given by the game engine. The goal is to develop a method that correctly identifies (clusters) these noisy versions of the same description. A DRRN baseline is employed and to improve its performance, the model is enriched with pretrained-embeddings. Performance is tested on 3 games generated by TextWorld. To further measure how the algorithms handle the stochastic information, cluster analysis case studies are done on the hidden states generated by the trained agents on the full database of descriptions.