Computational Linguistics vs. NLP

Have you ever talked to a scientist about their sciencing? They are geeks. They will ramble on, and on, and on about their field of study and never shut up. They will write entire blog posts about their sciencing! Kind of like what happens when you talk to an engineer—the proverbial wrestling in the mud with a pig. Here is my question: which is worse, the scientist or the engineer?

More seriously, this very question—"what's the difference between engineering and science?"—is in fact rather important in today's research on language and computation. One place where that is very clear is the peer reviewing guidelines of the major computational linguistics conferences sponsored by the ACL. Let me start by explaining away this jargon first; I'll get to my point in a minute.

The ACL is the Association for Computational Linguistics, the major scientific society in my field of study. Peer review is how we decide as a community of researchers if a paper looks scientifically sound and intellectually interesting; it basically consist in what your teacher used to do when they were too lazy to go over your reports and had students grade each other's work instead. Concretely, we get a bunch of submitted papers to read, and for each paper we fill in a questionaire with questions such as "is this paper appropriate for the venue it has been submitted to?" or "how interesting and impactful do you think this piece of research is, on a scale of 1 to 5?". There are three to five different anonymous reviewers per submission; some venues also include an author response phase to allow for clarifications where needed.

Peer review guidelines are a good indicator of what a field of study values: they outline very neatly the sort of elements that matter when deciding whether some submission is worth presenting at a venue. The reviewing guidelines of conferences like NAACL, EACL or EMNLP (all of which are linked to ACL), generally mention what "Prof. Philip Resnik at University of Maryland has to say regarding different kinds of contributions". In line with this long tradition of copy pasta, let me copy and paste the relevant citation, which I grabbed from these guidelines:

I think there would be significant value in encouraging reviewers to think explicitly about the nature of the contribution, and what questions then need to be asked. As a first pass for consideration/discussion:

Work in computational linguistics might include a mixture of scientific, engineering, and theoretical contributions, rather than just one. But, I am suggesting, if a paper does not make a contribution in any of those three categories, with the sub-bullets having understandable answers, one should seriously consider whether it belongs at the conference.

I found Philip Resnik's three-way distinction rather useful when approaching how I present my own work. I will confess that I am in general less passionate about the engineering aspects—I'm less interested in "problems to solve", and more attracted by "questions to raise". Having said that, I understand that it doesn't necessarily float other researchers' boat, as they may be more interested in clear-cut answers and definitive proofs. Trying to reframe my work so that it does a little bit of both has been a fun exercise.

Resnik also echoes here the distinction between "Computational Linguistics" and "Natural Language Processing" that I had briefly alluded to in a previous post. Here's an over-simplified way to explain that distinction: CL is more interested in the scientific aspect of the field, whereas NLP is more invested in the engineering aspect of it. What matters in CL is what your model explains, whereas what matters in NLP is whether your model works.

Let me stress it again: this is a gross over-simplification. A more subtle way to put it could be to say that NLP is the sub-domain of Machine Learning interested in language applications, whereas CL is the sub-domain of linguistics interested in computational models of language. And as per usual with scientists and engineers, if you put two in a room, you'll get three different opinions in no time, so this description of the distinction is probably not unanimous in the community.

Case in point: this second description, with CL being linguistics and NLP being Machine Learning, seems to me like it misses a point. CL works are generally relevant to the NLP community, and vice-versa. Sure, a CL paper might need a few touch-ups in order to use the jargon appropriate to the NLP subfield, but it's not as dramatic as visual aesthetic studies and the physiology of vision: although both field study vision and its effects, they are probably not highly mutually intelligible. In contrast, results from either CL or NLP are generally applicable to both NLP and CL.

Lastly, I should mention that there are some differences between the two, especially when it comes to the tools researchers from both fields employ. NLP has been dominated by artificial neural networks in the last few years. CL, and more largely speaking computational approaches to linguistics, rarely use deep neural networks—generally, they rely more on good old statistical tests and statistical models. I promise I'll get more into the weeds in upcoming installments to make all that clearer.

Now if you don't mind, I have some other things I must attend to. There's the author response due for some paper I submitted, and I have a few things to say about reviewer #2's remarks.