Here are some tidbits on academic writing that I originally posted on Facebook.

The irony of footnotes

Here’s the original post on Facebook, with many comments

The irony of writing footnotes in technical papers is that this very minor point that you meant to shove to the side becomes, visually, one of the first things to pop out when I skim the paper. I try to avoid writing footnotes whenever possible.


Here’s the original post on Facebook, with a lot more follow-up discussion. A few native Chinese speakers said this was helpful

I can always tell when an anonymous paper in review was written by a native Chinese speaker, if it frequently and inappropriately uses the word “Actually”. That word does not mean what the dictionary translation suggests.

Avoid the word “semantic” most of the time

Here’s the original post on Facebook, with comments. Maybe this is a lost cause.

One bit of sloppy writing that has permeated the computer graphics and vision literature is the use of the word “semantic.” Here’s why I think that you should avoid using it, or, at least, use it very carefully.

“Semantic” is a pretentious weasel-word. The word “semantic” is used in a way that, ironically, means almost nothing. However, it sounds like it’s implying some sort of insight about AI or human intelligence. I think that researchers use it when they want to indicate that there’s some high-level knowledge or context involved, but they’re too lazy to be concrete about it.

Instead of using the word “semantic,” I suggest thinking more concretely about what you really mean, and saying that instead. You will probably find that your paper is clearer. It’s a bit of extra work, but clear writing takes work.

As an example, our SIGGRAPH 2010 paper applied learned “semantic labeling” to 3D surfaces. I insisted that we avoid using the word “semantic” as much as possible. Instead, we wrote that our method learns to label object parts, such as “hand” or “wheel”, and that the labels can be chosen by a user. Saying that we learn to apply these labels is much clearer than saying that our labeling is “semantic” or that we label “semantic parts”, whatever that means.

Other typical uses (I am making these examples up) is to say “We let the user group regions based on semantic concerns” or “The video can be broken into parts based on semantics.” What do these sentences add?

Doug DeCarlo first pointed this issue out to me about a decade ago. He pointed out that “semantics” is the study of meaning, like dictionary definitions; how the phrase “I like it” means something different from “It likes me.” Objects in images do not have meanings in the same way. A hand or wheel does not have a meaning. He said that reading this usage of “semantic” felt to him like “nails on a chalkboard,” and now I feel that way too. (Images have meanings in a cultural context, but not in ways that computer graphics or vision normally address.)

There’s a descriptivist argument one can make: our community’s language naturally evolves over time. However, this doesn’t license arbitrary misuse of language; we shouldn’t use “plus” to mean “minus”. Misusing technical terms from other fields can cause lots of problems. Doug said that using “semantic” in this way makes you sound stupid to, say, a computational linguist who might review your grant proposal.

Whenever I see the word “semantic,” I think that the author hasn’t thought carefully about what they mean, and is only using this pretentious word because they think it sounds cool. Avoid being that person.

I do make one descriptivist exception at present: for better or for worse, the term “semantic labeling” has come to mean a specific task in vision and in graphics. So I think it’s fair to use the term in this case: this is the name of a task, and one must use shortcuts in names. The problem is that the word “semantic” is used all over in lots of other contexts where it means very little.

Update: Doug also points out this highly-relevant position paper by a pair of computational linguists.