What this seminar is about
This seminar is about methods of Bayesian inference applied to NLP problems. Our objectives will be for you to understand:
- what Bayesian inference is,
- how it differs from other probabilistic inference and modeling methods,
- how to intelligently incorporate uncertainties on one side, and domain-specific knowledge on the other, within a joint Bayesian model.
- how to use linguistic data to perform inference and solve specific tasks.
- how to assess and interpret the fitness of a learned model in the Bayesian setting.
To this aim, we will review a set of core Bayesian techniques through papers detailing their use in various NLP tasks.
General organisation of the seminar
The first 4 sessions will be dedicated to aquire the basic notions needed to understand research papers in the topic, and will therefore consists in standard lectures. The remaining sessions will consist in paper presentations by the students, followed by questions and discussions.
A working knowledge of basic probability theory is a mandatory requirement to be able to follow this seminar.
Everyone will be expected to have read the weekly discussed paper before every presentation, and everyone will be expected to give at least one presentation in the semester. Students taking the seminar for full credit will be asked to complete an additional assignment during the semester break. This assignment can either consist in writting a comparative study between a research paper that has not been discussed in the seminar and relevant litterature, or in a programming assigment and report in which student will be proposing their own NLP task and Bayesian model for this task, and implement an inference mechanism to solve it.
Precisions on this topic will be given at length during the first session of the seminar.