Meet the Researcher: Dominik Stammbach

Pictured: Dominik Stammbach (left); Jason Persaud ’27 (right)

Dominik Stammbach is a postdoctoral researcher at the Princeton Center for Information Technology Policy. Stammbach completed his PhD at ETH Zürich in Switzerland and is now a part of Professor Peter Henderson’s POLARIS (Princeton Language+Law, Artificial Intelligence, & Society) Lab, which conducts interdisciplinary research at the intersection of artificial intelligence (AI) and law. 

Stammbach recently sat down with undergraduate student Jason Persaud ‘27 to reflect on how his own interests developed as a student, why detecting misinformation can be tricky, and how he hopes his work will make an impact on justice. 

Jason Persaud (JP): Could you tell us a little about yourself and your work here at the CITP?

Dominik Stammbach (DS): Hello, I’m Dominik Stammbach and I’m a postdoc at CITP. I work on natural language processing and AI – mostly natural language processing, which has an impact on the world, public agencies, and public goods. 

JP: What led you to specialize in NLP [natural language processing], and how did your academic background shape your interest in this?

DS: Oh, that’s a good one. At first, I started studying in 2014 and NLP was sort of a random choice. On one hand, I was interested in computer science. I always felt like I’m quite good with computers. On the other hand, I was also interested in the humanities, and I was not quite sure about where I should go with studying humanities or computer science, and back then, it seemed like natural language processing sort of combined a bit of both.

And then I started studying that. It turns out that it’s been a very exciting 12 years in natural language processing since I started. And the better models get—the more we could certainly do with that kind of technology—the more intrigued I got. And at some point, I also figured out things seem to be working very well now for certain tasks.

I’m interested in how we can make these things facilitate work / assist workers. We can automate some things, which have high tolerance for making errors. So I don’t think we should try to automate things that are critical, but we need to determine: how can we use all of that and deploy it [technology] in a responsible way, and what are the good use cases for that? 

JP: Very nice. After looking at your background, a question that came to mind was – what do you see as the biggest challenges in applying NLP to areas like misinformation detection and access to justice?

DS: In a way, to me, the biggest challenge is that the current state of NLP is surprisingly advanced in tons of different areas, but at the same time it’s surprisingly dull in others. But if you just look at NLP output, it always seems like it always looks similar. So, the problem is that it’s just not that trustworthy.

Every now and then, things make mistakes. It’s impossible to get rid of those. But because you can’t really distinguish when to trust NLP output and when not, this brings a whole slew of challenges, which I don’t think will go away anytime soon. But then I think the question is not how to fix those things, but more about how we can use NLP technology despite knowing that we won’t have 100% accuracy on something. And I think there’s tons of use cases for that, nevertheless.

JP: Could you elaborate on some of those use cases? 

DS: Sure. So what I’m currently working on is legal natural language processing and NLP tech, which would increase access to justice.

In the US, unfortunately, most people cannot afford a lawyer. If you cannot afford a lawyer, chances are that things will not be great going forward. One thing we are thinking about is deploying technology for legal clinics or public defenders or other public agencies, which would make their work faster. One of the things people working in these offices is that they have to always retrieve relevant precedent and make something with that.

We think if you could automate finding relevant precedent, then this would make people way more efficient. Right now they’re searching for these things manually more often than not, with keywords and so on. 

We just think that having better search systems reduces the human time for which we do those things. And therefore they have capacity to work on other things. I think this sort of search is a good example here, because the worst thing which can happen is that you don’t return a result, but then humans have the agency to, you know, just re-run the search with a different input.

Societal Impact

So obviously your research is very interdisciplinary, especially focusing a lot on the societal implications of NLP. What impact do you hope your research will have on public policy and society at large? 

I hope my impact is indirect.

I’d love to do some research, which to some extent, strengthens public institutions, public agencies, and makes them more efficient or just slightly more efficient with NLP. And that, I think, in turn, has societal impact. For policy recommendations, perhaps I don’t have such strong views. That is not at the core of my research interests.

So I sort of think a direct impact I could see coming from my research is identifying what kind of NLP technology would be most useful to advance for public institutions. What is missing such that other people can make progress there? I’m not saying that I’m right.

I’m deploying the best possible legal retrieval system. But if I can sort of come up with: What is the goal? What is required by agencies? What are some of the things / features such as this “must-have?” And then I could come up with recommendations, perhaps even data sets for other people to work with and this eventually leads in aggregate to something which benefits public agencies.

JP: What piece of advice do you have for students interested in this kind of topic — this sort of intersection between legal studies and NLP? Any advice for students interested in either of the two, or the intersection of them?

DS: Yes. I once was in a class — Introduction to NLP. And then our lecturer sort of said that NLP is scary because it’s a combination of software engineering, advanced mathematics, optimization, linguistics, and statistics — and understanding all of that in detail is impossible.

But we actually don’t need that because there are lots of tools which abstract away from the query details. Now, if you say my interests are in NLP and humanities or society, then you don’t only have the NLP, which is a combination of lots of different areas, but you also have the societal bit, which again, can be very overwhelming.

But then, I think being brave enough to just have a high level picture of how things work and go forward with that seems reasonable. And where people can get lost is if they try to be on top of everything. And I just don’t think that’s possible anymore, in a world which gets more and more complicated.

---

Jason Persaud is a Princeton University sophomore majoring in Operations Research & Financial Engineering (ORFE), pursuing minors in Finance and Machine Learning & Statistics. He works at the Center for Information Technology Policy as a Student Associate, along with fellow undergrad Tsion Kergo. Together they are piloting a new series called Meet the Researcher.

The post Meet the Researcher: Dominik Stammbach appeared first on CITP Blog.