Interactive Question Answering

Artificial intelligence has long aspired to create systems that can perform real-world activities and converse with humans using natural language. In the pursuit of this goal, Interactive Question Answering was proposed. Specifically for language comprehension, IQA within text-based environments is seen as a viable direction to train and evaluate these systems. Three approaches are investigated: 1. the effectiveness of a more human-like policy-based reinforcement learning approach and an environment dynamics model to promote generalisation in textual environments. We evaluate the policy-based agent performance and the use of an environment dynamics model on the QAit (Question Answering with interactive text) benchmark. The results produced indicate that policy-based reinforcement learning has significantly better generalisation capabilities than the value-based QAit baselines and that an environment dynamics model can be used to regularise and promote generalisation when access to multiple training environments is not possible. 2. the use of Graph Attention Networks in providing additional context to agents to increase their question-answering abilities. Results indicate that Graph Attention Networks can increase the question answering accuracy of an agent when training on data containing high variability, thus improving the agent’s generalisation abilities. 3. framing an IQA trajectory as a sequence modelling problem. Using the novel Decision Transformer architecture, we investigate the applicability of a Transformer architecture in modelling IQA problems at scale. By utilising a causally masked GPT-2 Transformer for action generation and a BERT model for answer prediction, we show the Decision Transformer achieves performance greater than or equal to current state-of-the-art RL baselines on the QAit task in a more sample efficient manner.