Messages
A preliminary version of the WNNLP 2020 proceedings (with names of award recipients left out for the sake of suspense:) is available from here.
We’re happy to invite you to participate at the 2nd IN5550 Teaching Workshop on Neural Natural Language Processing, tomorrow (Wednesday 27/5) at 12:15. Please check Piazza for the Zoom invitation.
The four students enrolled in the class at the PhD level will be presenting their research projects from the home exam. The program is as follows:
12:15–12:20 Welcome and Workshop Overview
12:20–12:35 Ping-Han Hsieh: Targeted Sentiment Analysis for Norwegian Language
12:40–12:55 Maja Buljan: A Replication Study in Negation Scope Resolution
13:00–13:15 Awadelrahman Mohamedelsadig Ali Ahmed: Sequence to Sequence Learning for Named Entity Recognition
13:20–13:35 Ole Magnus Holter: Toward multilingual Named Entity Recognition for Norwegian and English
13:40–13:50 Award(s) Ceremony
Here's the final leaderboard for oblig 3:
1. janafag - 0.96
2. oysska - 0.94
3. zhenyinw, olemholt - 0.92 (tied)
Good job! You each get one extra point. :)
Tomorrow's (today's?) group session will be on using large pretrained architectures for building models, specifically with the transformers toolkit, something some of you had expressed interest in experimenting with for your exams. Same zoom link as usual :)
We have yet to publish detailed feedback in devilry (which, due to illness, may take a little longer this week), but we are happy to report that all students who submitted all three obligatory assignments are qualified for the final exam. There were 19 MSc and four doctoral students who submitted the first assignment, and everyone stuck it out with us this term, i.e. submitted all the way through the third assignment and, ended up qualified for the exam.
In light of the pandemic-related challenges to focused and productive studies, we lowered the minimum passing threshold for the MSc version of the class (IN5550) to 14 points, but in the end it turned out that this accommodation was irrelevant for the vast majority of students.
In the lecture tomorrow (Tuesday, April 21, 12:15, on Zoom), we will review the procedures and schedule for the final exam, with the three-week exam period starting tomorrow. As part of the lecture, we will provide a high-level overview of three common 'tasks' (i.e. sub-problems) in natural language processing, viz. named entity recognition, negation resolution, and sentiment analysis. Following the lecture, we will make available high-level project outlines for each task (e.g. indicating relevant background readings, research questions, and candidate approaches, data sets, and tools). Participants need to pick one of the tasks and declare their choice by the end of the day of April 22. Team submissions (of up to three MSc students; doctoral students are lone wolves) possible...
There will be a panic group session on the day of the deadline, 15.4.2020. Same time as usual, same link as usual (https://uio.zoom.us/j/5108530367). We'll be answering whatever questions you have about the obligatory. See you there!
The links for the zoom sessions for today's lecture are now available from the lecture schedule. Note that there are separate links for parts one and two.
We now publish the list of submitted systems which were the best in the Obligatory assignment 2, when evaluated on the held-out test set....
Earlier today, we have posted Zoom meeting links and slides for our first on-line lecture tomorrow; please see the schedule page for details. The lecture will be broken down into roughly two 45-minute segments, with a quarter-hour break inbetween; there are separate Zoom meetings (and links) for the two parts. There will be opportunity to ask questions and, more generally, interact with the lecturers and fellow students via Zoom. The lectures will be recorded and made available for download afterwards; please note that participation in the lecture establishes consent to recording, including of interactive student contributions.
Dear IN5550 students,
Now, as you are working on the obligatory assignment 3, we remind that you don't need GPUs for your IN5550 assignments. It is technically possible for you to run jobs on Saga GPU nodes, but this is not really necessary: especially given that GPU-accelerated nodes are scarce resource and must be consumed with great care. Many other UiO researchers are waiting in queues to use them.
The IN5550 datasets are small enough for training to be done on regular CPU nodes (and Saga has lots of them). Thus, in general we ask you to not submit GPU jobs in relation to your IN5550 class.
Thank you for your understanding. If you still have any questions concerning Saga or anything, we will always be happy to answer them via...
We had tried to avoid the proprietary Piazza environment this term, in part because UiO is expected to ban its use in teaching later this year. But our sense is that the Microsoft GitHub Issues tracker, which we have tried on as a replacement discussion forum for IN5550 this term, does not satisfy all communication needs around the class; notably, several of you have expressed a desire to be able to post questions anonymously. For these reasons, we have now (re-)activated Piazza as our primary support forum for the class and have sent invitations to all students earlier today (23 of you submitted the second obligatory assignment).
We are beginning to see more clearly about the path ahead: For IN5550, we expect it will be comparatively 'easy' to hold laboratories and lectures on-line, as well as to complete obligatory assignments and the final projects remotely. We have already lightened the lecture schedule and assignment load and expect to stick to the original schedule for the project-based exam. Working in groups remains a viable option (we think), though just now groups should probably not expect to work together through in-person meetings.
Assuming everyone has Internet access at home, our primary communication platform for lectures and laboratories (including interactive elements and individual consultations) will be zoom, which is available for all common operating systems, including on smart phones and tables. The university provides a...
We have just posted a revised schedule and update regarding the final obligatory assignment, to ease the transition to an exclusively on-line university. In a nutshell, there will be no lecture or laboratory session this coming week (on March 17 and 18), and we have reduced the requirements for the third assignment and extended its deadline.
We sincerely hope that everyone is well, given circumstances, and will be able to use the extra time to prepare for virtual interactions with us and fellow students! Starting on Tuesday, March 24, we will provide lectures and laboratory sessions in our usual scheduled slots; We will be presenting via the zoom platform and will inter...
The course teachers are getting together (virtually) tomorrow to decide on how best to move on-line the upcoming lectures in IN5550, and maybe laboratories too. We believe we should be able to largely follow our plans and schedule, but expect there will be far fewer (if any) in-person interactions in the weeks to come. We hope that everybody will be able to adjust their work environment and schedule to distant teaching, and we will be happy to hear from you about course-related suggestions or concerns. Expect more updates from our side towards the end of the day tomorrow, Friday, March 13!
We've published oblig. #3 on Github. Have fun!
The submit mechanism has been updated for oblig 2 -- if your models are too large to fit on Github, simply load the same module you use for PyTorch (nlpl-in5550/202002/3.7), and use:...
As promised, here is the leaderboard for obligatory 1 - all of you get 1 extra point (if you don't already have 6 points). Congratulations!
- oysska - 0.53
- diegog - 0.52
- majabu, pettemae, muhaaa - 0.51
We've published obligatory assignment #2 on Github. Feel free to start working on it, at least the first part - we'll discuss the rest in the group session.
For those of you working on the obligatory assignment 1:
Some of you can have problems with saving your trained models on Saga (usually it is accompanied by warnings like 'Couldn't retrieve source code for container of type...').
The reason is a bug in PyTorch 1.3. In case you encounter this, try loading the nlpl-in5550/202002/3.7 module instead of nlpl-in5550/202001/3.7. This version of the IN5550 Saga environment features PyTorch 1.4 where this bug was fixed. If we do not find any other regressions introduced in this release, we will make 1.4 the default PyTorch version for the rest of the course.
Good luck with the assignment!
This is a just a reminder that there is no lecture on Teusday this week, while the laboratory session on Wednesday will of course run at its usual time. The submission deadline for our first obligatory assignment is this coming Friday, so please use the opportunity to devote the extra time to polish your submission ... all points earned now will remain valid throughout the semester. In the meantime, most of us are attending neural teacher training in the Norwegian mountains these days ...
The first obligatory assignment is now published. You are free to start working on it. However (especially if you lack the machine learning background) it is advised to wait until the next lecture on January 21 and the subsequent group session on January 22.
The first lecture in IN5550 this term will be this coming Tuesday, January 15, at 12:15. We will go through course logistics (including routines for assignments and the final project-based exam) and motivate the now dominant use of neural architectures in Natural Language Processing (and most other sub-fields of Artificial Intelligence).
To prepare for this class, we ask that everyone fill in a brief anonymous survey about their background and request access to the Saga national supercluster.
There will be no screencasting in this class, and we will rather put our effort into making the lectures worthwhile, interactive, and fun. Both for obligatory assignments and the final project we encourage group work (of two to three students), and regular lecture and laboratory attendance will be a prerequisite to forming strong groups.