NLP Concepts Seminar: A Retrospective

teaching nlp by Vagrant Gautam

How the class I designed (on the conceptualization and operationalization of abstract concepts in natural language processing) actually went!


The summer semester is nearly over, and that means my seminar on defining and measuring abstract concepts in NLP is also almost done! The class exceeded my expectations: My students were very active in discussions, bringing very different perspectives to class from their personal, professional, and academic experiences, and we all learned a lot. And according to the course evaluations, the students seemed to really enjoy this class, in spite of the fact that I assigned four (4!) times as much reading as most other seminars here.

I previously posted my concepts syllabus; in short, it's a class I designed to explore abstract concepts such as interpretability, bias, etc., in terms of:

  1. How they are currently conceptualized and operationalized (i.e., defined and measured) in the field -- by reading 3 popular/recent papers per concept
  2. What these choices might miss -- by reading 1 interdisciplinary critique / conceptual paper

After some general reading on conceptualization, operationalization, and how to read papers, the concepts we ended up covering were: Names, interpretability, reasoning, stereotypes, paraphrases, emergent abilities, explainability, and bias. I presented names, interpretability, and stereotypes, and the rest were paired student presentations.

What went well

Assignments

Before each session, I had students submit a short (3-5 bullet points) graded assignment reflecting on the four readings. I explicitly allowed them to write more if they were feeling particularly inspired, and I consistently had many students take me up on this right from the start. In my instructions, I asked students to synthesize ideas across all the papers and critique the concept papers or point out what the authors did especially well. Some students had trouble with this at the beginning, instead just summarizing what each paper did (e.g., some students went too low-level with their critique, so I gave them feedback to think of bigger, more fundamental problems with how a concept was being treated in a paper). But by the end of the term, all of my students were submitting solid assignments: Sometimes they brought up points that I know other people have written about, and I enjoyed sending them paper recommendations in my responses to their assignments. They also made many connections in these assignments that I hadn't thought of and that were really fun for me to engage with.

Discussion

Since it was a seminar, I wanted the class to be discussion-heavy. I also wanted the discussion to be high quality, and the assignments were my approach to achieve this, encouraging students to do the readings even if they weren't presenting. Overall, I was very happy with the quantity and quality of the discussion -- students made connections across sessions and concepts, and connected their own experiences (personal and professional) to the topics we were discussing. Discussion-heavy classes can sometimes be challenging for students who are shy or less confident in themselves, so I tried to ensure that everyone got a chance to speak. I also explicitly told student presenters that they would be graded on their classroom management and on allowing everyone to participate in discussions, because I believe this is an important skill, regardless of whether it's in a classroom in an academic setting or running a meeting in industry. This worked exceedingly well, and a third of my student evaluations explicitly mentioned it ("Welcoming atmosphere, easy to participate in discussion," "[The lecturer] always made sure that everybody was included in the discussions," "Learning atmosphere during our sessions allowed me to speak freely and engage in discussions which is sometimes difficult for me"), which I was thrilled to hear.

Cohesion of the readings

Interpretability, explainability, and reasoning were concepts that worked especially well, in my opinion. Some concepts also overlapped (e.g., interpretability and explainability, as well as names, bias, and stereotypes), which was really nice for making connections across sessions and concepts. For explainability, we read a critique of AI explainability that surveyed social science papers on explanation from various disciplines, which helped ground our understanding of explanations in the other papers. We also read 'Attention is not Explanation' and 'Attention is not not Explanation', which worked well together because the latter is a response to the first, and itself a critique. Interpretability was another concept with two critique papers, 'The Mythos of Model Interpretability' and 'Against Interpretability', which provided great perspectives to ground our reading of Patchscopes and RAVEL, two rather technical papers. Reasoning, on the other hand, didn't have an explicit critique paper, but it still worked well because we covered a range of types of reasoning (pronominal, spatial, logical, mathematical, etc.), as well as recent approaches in the literature (chain-of-thought prompting and DeepSeek's reasoning chains). This diversity gave us a good bird's-eye view of work on reasoning and helped us find similarities, differences, and missing pieces.

Pain points

Connection between assignments and sessions

The assignments and sessions weren't well connected in a few different ways. First, the student presenters never had access to their peers' assignments, because I wanted students to be able to do their readings and submit their assignments right before class instead of significantly earlier. Secondly, this also sometimes meant that some really great points from the assignments were never discussed in class, just because they happened not to fit with the discussion prompts the presenters came up with. Since I had access to all the assignments as the instructor, I was much more aware of this gap than the students were, but at least one student did tell me that they felt there sometimes wasn't space in the sessions to discuss thoughts they had written down in their assignments. So next time, I want to try to have a tighter connection between the assignments and the sessions -- perhaps by moving the assignment deadlines up to incorporate parts of them into the session, or by having an open-ended component to the sessions where students could bring up observations we hadn't already addressed.

Cohesion of the readings

There were a couple of papers that I found simply boring or a slog to read, which I would remove next time, and one conceptual paper (What is a paraphrase?) that should either be replaced with a critique paper or removed altogether. At a higher level, since I prioritized student preferences for presentation dates, concepts that are closely connected ended up being far away from each other on the schedule. This made it hard to disentangle the significant overlap between interpretability and explainability, for example. Next time, I'd love to combine related topics, extend the time we spend on them, and add the following readings which I mentioned in class and would provide excellent additional context to our discussions:

Workload

According to the course evaluations, roughly half the students thought the requirements and workload were just right, and half thought it was a little demanding. This was something I was concerned about, but since student satisfaction was so high on the overall metrics and since not a single student selected the "too demanding" choice in any of these questions, I'm not sure whether I should actually make any changes here. I'm reluctant to assign less than 4 papers per concept because I still feel that it's necessary for the kind of high-level discussions I wanted us to have about a concept, so perhaps having fewer concepts and spreading them out over a longer period would be something to consider. As for my workload, the individualized and extensive feedback I was giving was fun but far too time-consuming to scale to a group larger than 10 students. I struggled to find time to do it amid submissions and other work, until after submitting my thesis.[5] So if I wanted to run a class like this on a larger scale, I would need to restructure the class as well as my feedback.

Summary

Overall, I think the seminar went swimmingly. The students who did all the readings read 34 papers and got a new appreciation for some pressing problems in NLP. They turned into skeptics and good conceptual thinkers by the end of the course, paying attention to both the social and technical aspects of NLP, and how decisions about conceptualization and operationalization can impact all of our science as well as public perceptions of NLP technology and tools. At the beginning of the course, I told them I would feel like I had succeeded at running this seminar if I could get them to come up with new ideas that I hadn't thought about or read about before; they did this over and over again at every session, and I'm so proud of them!

  1. Tim Miller, Piers Howe, Liz Sonenberg ↩︎

  2. Naomi Saphra, Sarah Wiegreffe ↩︎

  3. Jasmijn Bastings, Katja Filippova ↩︎

  4. Sarthak Jain, Sarah Wiegreffe, Yuval Pinter, Byron C. Wallace ↩︎

  5. Oh, yeah, I submitted my thesis! ↩︎