17 January. 2021
When a human interacts with their machine assistant, the scope and language of their interaction is typically fixed by the developers of the assistant. Even in those cases where the assistant is capable of adapting to the human, the feedback that the human provides to facilitate this process is typically very restricted, and comes in the form of categorizing objects into classes (in support of supervised learning) or reacting to the execution of policies (in support of reinforcement learning). Yet, in real life, the adaptation of a human assistant comes through a much richer form of interaction, where both the assistant and their counterpart engage in a natural language dialogue, with each interlocutor explaining why they chose to take an action and/or why they consider an action to be (in)appropriate. This process of coaching, rather than simple supervision, allows for the much more efficient and robust transfer of knowledge to the assistant. This project seeks to investigate how this process of coaching can be used when the assistant is a machine (see https://www.researchgate.net/publication/334989337_Machine_Coaching ), focusing on how the machine will generate natural language explanations for its internal symbolically-represented knowledge and inferences, and how the natural language explanations received by the human counterpart will be turned into this internal representation. Depending on the internship type, the project can be adapted towards the natural language processing or generation part, towards the formal analysis of the coaching process, or towards the empirical investigation of the effectiveness of a natural language interaction for machine coaching.
Natural language processing OR natural language generation OR formal analysis OR empirical cognitive psychology.
The main objective of the internship is to investigate how the process of coaching can be used when the assistant is a machine, focusing on how the machine will generate natural language explanations for its internal symbolically-represented knowledge and inferences, and how the natural language explanations received by the human counterpart will be turned into this internal representation. Depending on the background of the intern, the project can be adapted towards the natural language processing or generation part, towards the formal analysis of the coaching process, or towards the empirical investigation of the effectiveness of a natural language interaction for machine coaching.