I am a PhD student who started recently in June at Utrecht University, under the supervision of Dr. Daniel Oberski and Dr. Dong Nguyen. The focus of my PhD project is on improving measurements of and causal inferences with theoretical (latent) constructs (e.g. personalities, human values) by leveraging high-dimensional textual data.
Following my supervisors’ advice, I signed up for the ACL 2020 conference with two main goals: First, become more familiar with the ongoing research in computational linguistics (and also the language used in the field, as I come from a statistics background); Second, connect with other researchers, especially those who work on similar topics as I.
See below my conference week schedule:
Note: I didn’t attend all the sessions I planned, as is explained later in this post; the empty slots were meant for reading papers and going through pre-recorded materials.
First Day Tutorials and My Strategy
The first day of the conference was mainly about tutorials, each of which typically lasted several hours. I did not know in advance that I would need to watch the pre-recorded videos prior to the live tutorial sessions (my bad!), so when I logged onto the virtual ACL conference system on Sunday to attend a tutorial I signed up for, I was caught unprepared. Knowing that the pre-recorded videos are simply too long to watch for me to be ready for the discussion sessions, I found and installed a plug-in that would speed-up the videos, so that I could finish watching the pre-recorded videos sooner. However, since I was (and still am) new to the field, not all videos were easy and quick to digest for me. So in the end, I spent the whole day trying to catch up with all of the pre-recorded materials, leaving me little time to actually come up with questions and attend the discussion sessions.
Having had a less-than-ideal start and noticed how many more exciting events were yet to come in the rest of the conference week and that they would usually require some form of preparation, I realised that I should use a strategy that would allow me to spend my time and energy in a more selective, effective way. Take the main conference (Monday - Wednesday) for instance, It would be simply unrealistic to read all the interesting papers (there were many!) and attend all the respective sessions!
So, I made a list of papers I found to be interesting and related to my own project, sorted them into three separate days (of the main conference) and narrowed the list down to be about 5 - 7 papers a day. Then, I would go through the pre-recorded presentations of those 5 - 7 papers that were short-listed for the next day and make a final decision on ONLY TWO papers that I would actually read and whose sessions I would attend (with questions).
Apart from the paper sessions, I also decided to join a group mentoring session, a Birds-Of-A-Feather (BOF) meet-up and of course, the keynotes. This strategy turned out to work extremely well for me throughout the conference. Also note that despite only two papers a day, I still spent more than 10 hours a day at the conference, which is more than equivalent of a regular workday.
The Paper Sessions
I very much enjoyed the paper sessions. Normally I would feel quite nervous about initiating a discussion in an in-person meeting, especially with people I did not know before. I would think very carefully about what I want to say and when I would like to say it, just to avoid saying something embarrassing or rudely taking over someone else’s turn. Somehow this turned out to be much less of an issue at the virtual conference. I joined every paper session right on time when it was supposed to start and because of that, I often happened to be the first one there, which naturally led to a smile, a “hi” and a follow-up conversation about the paper. This “natural” flow of things prevented me from stressing out too much or over-thinking, which I appreciated. The extraordinary amount of friendliness and patience of the authors with whom I spoke helped too.
Furthermore, probably because I was usually the first to speak and that others were kind enough not to interrupt me, I was able to ask all the questions I prepared and engage in substantial, rewarding discussions with the authors. In this sense, I felt that this virtual conference experience was much more personal to me than expected (and I guess not necessarily less so than in an on-location conference).
After my part, I tended to stay a little longer in the sessions, curious about what others had to say about the papers. But sometimes I just thanked the authors and left the chat either to read another paper or to take a break, knowing it was completely fine to do so. I know that some conference attendants also switched between the paper sessions to maximise their experience. Such a flexible strategy might not have been possible if the conference were to take place in person.
There were even more surprises: You may just happen to be in a session with someone famous who you only know from the literature or online blogs, which completely happened to me. I also had a small friendly debate with someone who I did not know at the time to be a very big name (mainly because I don’t know the names in the field very well yet).
I decided to attend a BOF session (computational social science track). There were more than 45 participants, which a one-hour session cannot easily accommodate. Luckily, everybody got to introduce themselves and their work (thanks to the awesome host Diyi Yang too). Because of that, I met a colleague who works on a similar topic as mine and we got to speak with each other privately later. I definitely think that there could have been more social events like this, and that they can be preferably organised in smaller groups (for more meaningful interactions). Perhaps this can be considered for the next conference, be it virtual or not.
Group Mentoring Session
I attended a group mentoring session. Because of some technical problems with Zoom, some participants were kicked out (multiple times) and we lost quite some time just trying to wait for everyone to be back in the chat. This was unfortunate, but our mentors (Chris Brew, Liling Tan and Malihe Alikhani) stayed extremely helpful and patient and provided very good tips on doing a PhD. I appreciated it very much.
In between the paper and social sessions, I also watched the two very high-quality, inspiring keynotes from Prof. Kathleen R. McKeown and Prof. Josh Tenenbaum. I could not attend the workshop live sessions as I planned to because of some family emergencies. Having (happily) learned that most presentations and talks would still be available on the conference website (and later be made publicly available) after the conference week is over, I decided to postpone the pre-recorded presentations and invited talks and some papers I found very interesting to the week later.
My Main Impression
Overall, I very much enjoyed my very first conference. It was definitely an intense experience, as there was so much input to take in, so many decisions to make, papers to read, and people to get to know. I ended up spending more than 10 full hours just at the conference alone. But it was also very rewarding!
First, I received an up-to-date overview of the field of computational linguistics (e.g. what people work on; what are the most relevant topics these days; what language do people use; who does work that is similar to mine). Second, I read some really interesting papers and had very meaningful discussions with the authors. With a background in statistics and research methodology, I was very happy to notice in some papers that the principles of statistics and research methodology widely used in the social sciences are now being applied to computational linguistics. Third, I connected with some very nice researchers who work on a similar research topic. I can see much better now the value of my own research project, its connection to the field of computational linguistics and the challenges it is faced with. Lastly but not least, I was greatly inspired by the many talks I watched and the conversations I had. Among many others, I especially appreciated what Prof. Kathleen R. McKeown said that we should focus on research tasks that really matter and bring the language aspects back to NLP (e.g. examination of model outputs and careful preparation and analysis of data), rather than stress too much on only the state of the art performance on specific data sets.
Let me mention a few more things I particularly appreciated about the conference. First, the community. You can easily see in the live sessions and chat channels how approachable, appreciative and supportive many researchers and attendants are. This I believe to be very crucial for a healthy, encouraging and nourishing research environment (especially so for newcomers like me). Second, the pre-recorded videos. The obvious advantage compared to a live presentation is that you can pause and rewatch the presentation at your own pace. This is really helpful especially if a presentation deals with a complex topic that is just difficult to follow in a live session which you cannot pause or rewatch. In addition, a pre-recorded presentation guides you through the respective paper, which on the one hand helps you decide whether you want to invest more time into reading the full paper, and on the other hand, facilitates your understanding when you actually read the paper. Third, the flexibility. This is made possible partly due to the pre-recorded sessions that remain available even after the conference week, and partly thanks to the particular virtual conference set up (e.g. multiple identical sessions to accommodate different time zones; the Zoom meetings). If you want to, you can attend multiple parallel sessions by switching between the Zoom meetings and have all the discussions you would like; if you are the less energetic type, you can opt for short discussions in a small number of live sessions and leave right after without appearing awkward, so you can retain energy for the papers that you still want to read. If you are busy and have to miss a really interesting session, you still get to watch the pre-recorded presentation (and sometimes also the live session) later, thus not having to feel too bad about missing out on things.
Tips for New PhD students
Like every new, enthusiastic PhD student, I have the tendency to think that I need to learn and accomplish as much as possible, also when it comes to attending conferences. However, with many helpful insights from my peers and supervisors, I am slowly starting to realise that this is an unrealistic, even counter-productive goal to set.
Unfortunately (and fortunately), on a given day there is a limited amount of new information that we can process and retain, just a small number of new people we can initiate meaningful conversations and further relationships with, and only small steps of work progress that we can make. Aiming beyond these (realistic) limits will bring probably only marginal additional gains while leading to serious problems like work-life imbalance, health complaints, decreased work efficiency and lower life satisfaction in general. That is why I now try to set more realistic work goals for myself so that I can focus on a selective number of things that matter the most and if I manage them, I can be proud of myself. I applied this strategy to ACL 2020, where I decided to set achievable goals for myself after the somewhat chaotic first day, which worked out really well for me and I am very happy about what I took away from the conference.
So my advice to new PhD students like me when attending conferences (especially if it’s the first time) is to set minimal, realistic goals that are tailored to your very own situation. If you are new to research and cannot digest a paper quickly enough, then set for yourself a goal of just one paper a day. If you are very shy and feel intimated by interactions with more senior researchers, go for a session led by PhD students and ask just one question. If you think you need to practice your social skills, then perhaps consider initiating a private chat with a new colleague. Of course, always measure the goals against your own current abilities. It is always better to set achievable goals that are ideally within or just slightly outside your comfort zone. This way, you will more likely end up taking something pleasant and meaningful away from the conference experience and feel a sense of achievement.
Last but not least, many thanks to the conference organisers and participants for making a big, successful conference like ACL 2020 possible!
Below are some papers I very much enjoyed reading during the conference (in no particular order).
Shah, D., Schwartz, H. A., & Hovy, D. (2019). Predictive Biases in Natural Language Processing Models: A Conceptual Framework and Overview.
Ethayarajh, K. (2020). Is Your Classifier Actually Biased? Measuring Fairness under Uncertainty with Bernstein Bounds.
Keith, K. A., Jensen, D., & O’Connor, B. (2020). Text and Causal Inference: A Review of Using Text to Remove Confounding from Causal Estimates.
Joseph, K., & Morgan, J. H. (2020). When do Word Embeddings Accurately Reflect Surveys on our Beliefs About People?.
Ribeiro, M. T., Wu, T., Guestrin, C., & Singh, S. (2020). Beyond Accuracy: Behavioral Testing of NLP Models with CheckList.
Arora, S., May, A., Zhang, J., & Ré, C. (2020). Contextual Embeddings: When Are They Worth It?.
Gururangan, S., Marasović, A., Swayamdipta, S., Lo, K., Beltagy, I., Downey, D., & Smith, N. A. (2020). Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks.
Xia, M., Anastasopoulos, A., Xu, R., Yang, Y., & Neubig, G. (2020). Predicting Performance for Natural Language Processing Tasks.