Bootstrapped Q-learning with Context Relevant Observation Pruning to Generalize in Text-based Games