Student Training for Distractible Teachers

Peter Hollows · 2022 · Stanford XCS224U machine-learning
Download PDF

Discussed on the blog: The Literal Listener

Abstract

My project seeks to improve open-domain question generation. I introduce a student-teacher game played in the presence of distractors (i.e. a distractible teacher), in which a pragmatic speaker (the student) has to reason about a literal listener (the teacher). To score this game, the metrics of topicality and teacher patience are introduced, which are calculated from the teacher’s perspective. These metrics are used as a reward to improve pretrained QG models, and to evaluate different experimental approaches. I test the hypothesis that a topic embedding of the teacher’s knowledge better enables the student to generate questions that recover hidden knowledge in the game. I also test the hypothesis that reinforcement learning can be used to further increase the student’s ability to interrogate the teacher through additional training on a large, question-free corpus. Results are positive but limited; with hypotheses proving true by the numbers but falling short in qualitative assessment.

@misc{hollows2022studentt,
  author = {Hollows, Peter},
  title  = {{Student Training for Distractible Teachers}},
  year   = {2022},
  month  = mar,
  note   = {Stanford XCS224U},
  url    = {https://dojo7.com/papers/xcs224u/}
}