|20.04.2020 - 20.07.2020|
|Moodle page, course RL4NLP id 23952|
|madureiralasota @ uni-potsdam . de|
If you are considering taking this seminar, please write me an e-mail and register in the Moodle course, to help me plan our schedule. You can also tell me if you have any preferences regarding the papers in the bibliography, or suggest a paper that is not cited there yet.
Reinforcement Learning (RL) is a machine learning paradigm that is becoming more popular for Natural Language Processing (NLP). In this seminar, we will cover recent literature that apply RL to different NLP tasks, discussing how NLP problems can be modeled with RL concepts and comparing the results to other traditional or common approaches.
An important skill that all NLP researchers and practitioners must have is being able to critically read papers. Just because a paper has been presented at a conference or published on a journal, it does not necessarily mean that it is flawless. We, as responsible readers, must be able to understand the methodology, evaluate the approach and build upon the work of our colleagues to improve our field. Besides becoming up-to-date with the applications of RL in NLP, the purpose of this course is also helping students getting used to constructively assessing and discussing scientific literature.
Disclaimer: due to the format of the course and the limited time, this course will not teach Reinforcement Learning per se. The main abstract concepts will come up in the discussions, but if you wish to learn its mathematical foundations or algorithms in detail, take a look, e.g. at David Silver’s video-lectures, Emma Brunskill’s course and the reference book by Sutton & Barto.
This course has been adapted to happen 100% online due to the current pandemic, so it will be mostly reading/writing-based.
You will work on a scheme I named “bloggers for one week”. This idea derives from the ever-increasing importance of well-written blog posts and tutorials that help us in our learning process by demystifying and elucidating papers or methods that look difficult at first, especially due to the highly dense content the authors normally have to compress into a limited number of pages.
So each week one participant pretends to be a blogger and writes a critical/explanatory review in the form of a blog post about one of the papers in the bibliography. This will be posted on Moodle together with three topics you propose for group discussion in the forum with your peers. Throughout the week, everyone takes part on the debate and you mediate the forum. Then you and I meet virtually for a short spontaneous talk about your impressions, so that I have the chance to see you once during the semester and give you feedback. In the end of the semester, you will work on a review article or on a Python implementation as part of the final grade. See details on the guidelines.
Everyone is welcome to take this class, but you will profit more if you are already familiar with the main NLP tasks and methods. Watching the online courses I mention above is also a great help. If you are in doubt, contact me :)
To pass the course, you must achieve a performance of at least 60% on each of the items above separately, which includes abiding by the guidelines listed here. If you are a student of the CS department in Uni Potsdam, the minimal performance for the CPs to be valid in your course is 70%.
Feel free to write me if you have any question or suggestion. I hope to hear from you soon!