Towards Two-Way Interaction with Reading Machines

Sebastian Riedel, Sameer Singh, Guillaume Bouchard, Tim Rocktäschel, Vicente Iván Sánchez Carmona

Abstract: As machine learning models that underlie machine reading systems are becoming more complex, latent, and end-to-end, they are also becoming less interpretable and controllable. In times of rule-based systems users could interact with a system in a two-way fashion: injecting their own background knowledge into the system through explanations in the form of rules, and extracting explanations from the system in the form of derivations. It is not clear how this type of communication can be realized within more modern architectures. In this position paper we propose a research agenda that will (re-)enable this two-way communication with machine readers while maintaining the benefits of the models of today. In fact, we argue for a paradigm in which the machine reading system is an agent that communicates with us, learning from our examples and explanations, and providing us with explanations for its decisions we can use to debug and improve the agent further.