Natural Language and Artificial Intelligence

“Most of the tasks in which current AI systems excel are not natural tasks for humans. Activities such as playing games, recognising disease states from medical images, finding answers to queries from terabytes of text are activities which require specialist training for humans. Such AI systems make use of vast quantities of task-specific information, in a way which is impossible for humans. In the case of games, the AIs can create such information themselves, by playing incredible numbers of games while learning. In other cases, they can access huge quantities of data with great rapidity. AI systems are generally less impressive with activities which are more natural for humans, such as car driving. In this talk, I will focus on AI as it relates to human language, where machine learning techniques have allowed us to make huge advances recently, but where the state-of-the art is very far from the abilities which humans acquire naturally in childhood. I will discuss the extent to which insights from linguistics are reflected in current approaches, and whether we might make more rapid progress by looking at techniques which combine symbolic and neural approaches. I will illustrate the talk with examples from our current research in Cambridge.”