Richard Wicentowski, more commonly known as Rich, is a computer science professor. He introduces many students, including myself, to computer science. He shared insights on how he integrates text and computing systems in his work and courses.
Shinz Jo Ooi: What inspired you to go into computer science, or more specifically, computational linguistics and natural language processing (NLP)?
Rich Wicentowski: When I was seven or eight years old, my mom brought home a computer from an after-school program. I was born in the ’70s, and it was very uncommon to have a computer at home then. But, because she volunteered at this program, where students only used the computer on Saturday when the class was held, we got to bring it home and have the computer in our house for six days a week. I got to play with it all the time, so I got interested then.
Because of that experience, I’ve been into having a computer, or, a programming computer, ever since I was a kid. I’ve also had a computer that was my own since I was maybe ten years old, and that’s, again, at a time when it was somewhat uncommon for people to have computers in their house.
When I went to college, I was sure I was going to do computer science. But, after taking some linguistics classes, which I really enjoyed, I actually briefly contemplated not being a computer science major and being a linguistics major. At one point, I thought about an anthropology major, but in the end, I decided that I really liked computer science the most … I ended up doing computational linguistics because it brought together my interest in linguistics and my interest in computer science. I ended up not using anthropology in any way in my future work, although I guess it’s related in a sense by studying human language as an artifact of something that people have made.
SJO: How do you think you blend these two aspects together in your work now?
RW: I think on a daily basis, it’s not obvious that I blend them together, although I think my training means that the way that I think now and the way that I approach problems often comes from that background.
I think the kinds of problems that are interesting to me tend to be problems that involve text. Even though maybe my work now is less focused on linguistics, it’s still focused on text. The things that interest me are things where data is stored in some text format, and I want to do something interesting with that text data, even if the problem that I’m studying isn’t something about the morphology of French, for example, which would have been a linguistics problem that I actually did work on in my Ph.D thesis, but it’s not a problem that I would work on now.
SJO: Do you have any projects that you’re working on right now that you’re comfortable sharing?
RW: Sure. I have an article in submission that’s looking at how changes have been made to medical clinical trials over the past ten years. When you do a clinical trial, you have to file a report with the government to say “Here’s all the things that we plan on doing.” Examples of some of the parameters you could change when designing a clinical trial could be doctors knowing or not knowing who the control group is. You might test one drug against a control group, or you might test two different drugs against the same control group. There’s been a lot of work that has studied clinical trials to say, “What are the best practices? What are the best ways that you should design a clinical trial?”
The paper I have is examining whether or not, in the ten years worth of data from 2010 to 2020, people have been adopting the best practices for clinical trials based on the things that have been written that say “Here’s how you should run a clinical trial.” We have a data set of all of the clinical trials that have been submitted to the government, and the text of what they say they’re going to do. So, if in 2017, it came out that people should be using a particular technique in their clinical trials, do we see evidence that in 2018, 2019, 2020, there is this increased adoption of that technique? Or are people who run these large clinical trials basically ignoring the scientific literature and not adopting them? Or do you see adoption rates higher in academic research settings and lower in industrial research settings? So those are the kind of questions that this paper is trying to address, and it was actually just submitted earlier this week.
It’s the third paper that I have kind of along the same set of data, and I’ve been working on this project probably since 2018. In 2018, we were able to get the data for a single year’s worth of trials as a pilot study, and we showed that we were able to extract what we were interested in extracting. We had a paper published on that, and it was well received. And based on that, they gave us ten years worth of data.
That’s not linguistics for sure, but it uses techniques about extracting information from text data. A large portion of the clinical trial data is just text, where someone just types and says what they’re going to do. Having to read through that text and understand what it is they’re saying is why I was involved in the project.
SJO: What has been your favorite class you’ve taught at Swarthmore so far?
RW: My favorite class in terms of content is actually a course that we don’t teach anymore, which makes me a little sad. It was a course that originally came between CS 21 and CS 35: Data Structures and Algorithms, and much like CS 31: Intro to Computer Systems, you could have taken it before or after 35. But it was numbered CS 22 [Structure and Interpretation of Computer Programs], which was designed to come after 21 but many students didn’t take it until after they took 35.
I would say that I really enjoy teaching all my courses, but CS 21 is great, because you meet so many students across campus who you know won’t end up [taking] CS, but you get to expose them to something that they never saw before and get them, or some of them, excited about it. It’s not like we’re struggling to have majors in the department, and we’re recruiting students from that course, but I think it’s really important for us to make sure that students who didn’t know that CS was for them know that CS could be for them if that’s what they want to do.
I really like teaching CS 31. I think there’s a lot of really cool topics in that course, and there’s a lot of aspects to that course that are really interesting and fun. I also suspect I’m gonna have a blast teaching [Computing with Text Systems] next semester.
SJO: You’ve mentioned that tools like ChatGPT have been one of the huge changes in the NLP field. I was thinking about how these large language models might have affected the techniques you use in NLP and computational linguistics; does that hold true?
RW: So there’s a split between natural language processing and computational linguistics, and people use the words somewhat synonymously at times. The linguistics department here teaches a computational linguistics course that [Associate] Professor Jonathan Washington offers, and our course was always a natural language processing course. To think about the divisions between them, computational linguistics is about how you use computers to study language. Natural Language Processing is about how you use text and language to solve problems that might not be about language. For example, one task might be if someone hands you a Yelp review, and you don’t have the number of stars — can you write a program that figures out whether or not the person liked the restaurant? That’s not a linguistics question, but certainly it’s about understanding how to parse a sentence and what the words mean. If someone says “I did not like it,” that’s different from “I did like it,” right? The word “not” is important there, so understanding morphology, syntax and semantics is important for those problems, but you’re not studying how the semantics of this particular language works, which would be what computational linguistics does.
Thinking back to my own Ph.D research that was studying computational linguistics, it was studying how morphology worked across a large number of mostly Western European languages that used single letter alphabets, as opposed to something like Chinese that would use a character to represent a whole word or most of a word. So, it was studying how to understand the morphology of languages without actually any real human input. But the end goal of that wasn’t really to make deep insights into the language, which I would say is what a computational linguist would do. My work was more on thinking about how identifying the past tense of some verb in Spanish can improve the performance of machine translation later on as a downstream task. So although it was a linguistics task, it wasn’t really for studying the language.
Certainly, large language models have affected natural language processing quite a lot. As someone who’s not really as involved in computational linguistics, I don’t know how much, for example, someone like [Washington’s] work has been affected by large language models. That would be really interesting to find out. But [given that] large language models are being trained largely on English text or for other well-sourced languages like French and Chinese, there are not large language models developed for many of the languages that people are studying. If you’re someone who’s going out in the field to study some language that’s only spoken by 250 remaining speakers, you’re not building a large language model for that, so I don’t know how much it’s impacted their work.
In terms of natural language processing, the whole field is very different than it was ten years ago. Had you taken my course in 2018, there’s certainly a lot of it that you could still talk about, because the fundamentals of the field are still there, like understanding what words are like. I know that sounds silly, but understanding how to take a piece of text and break it up into individual sentences is important. Language is not just this giant hunk of text, there’s structure in it that’s useful for us to understand before we put it into a large language model. Still, I would say that a large part of the class, like semantics, machine translation, summarization, are completely different now as there are large language models, versus what people used to do ten years ago.
For example, in summarization, a very common thing to do if someone in natural language processing wanted to summarize a newspaper story is they would read the newspaper story and then try to find the two or three most important sentences in it, and then return those two or three sentences to you as the summary. Now, of course, that’s not what a large language model would do. It would actually “read” the story and then return to you a summary of the text in different words than in the original story.
It’s obviously very different than it used to be. It’s a challenging course to think about teaching, because the experience that you need to teach and to take that course is to have some machine learning background first, [since it is] a course focused on deep learning applied to natural language processing. Our department has shied away from having prerequisites in our upper-level courses, so it would be a pretty big change for us to start offering courses that you can only take after taking another upper level course first. The Computing with Text course that I’m teaching next semester allows me to teach pieces that are still relevant from the natural language processing course, but also teach the things that I think are really interesting about what I do.
SJO: One final question: what’s the greatest pride of your career so far?
RW: That’s a super hard question. I’m really proud of my students when they go on and do great things. I know that I had a small part in that, even if it’s a student who took my class and didn’t do great and really struggled, but got something out of it, and went on a journey besides computer science. I think having an impact on student lives is really important. I also think that I feel like I’ve had an important impact on how the college runs, and that’s important to me. There’s a lot of things that I’m really happy about how things have turned out at Swarthmore, and so I don’t know if I have a favorite.
SJO: Thank you for doing this interview with the Phoenix!
