TODO: Write something about this.
TODO: Write something about this.
TODO: Write something about this.
Many [1] of my students like to write code more than they like to write latex (i.e., the paper itself) [2]. Obviously, we desire everything: novel contributions, broad claims that are empirically supported in a convincing manner, insightful analyses ‐ and everything well explained in fluent, typo-free prose. However, as deadlines come close it is necessary to triage. Otherwise, we end up with something like this, as someone has conveyed on Twitter.
My advice:
- Well-written good results are better than poorly-written amazing results. Stop working on yet another model or technique!
- Strong empirical support for a narrow (perhaps less interesting or novel) contribution is better than unconvincing support for a broad (or interesting, insight, surprising, etc.) contribution.
Write early!
In fact, I suggest this heuristic, even long before any paper deadline: ask your self, "Will this be part of the final paper?" If so, then it lies on the critical path, so you'll have to write it up sooner or later. Sooner is usually better.
The corollary is that you should write up baseline results as soon as they are complete. By "write up" I mean: describe the baseline approach, including data, experimental setup, and results. At this point, you probably already have an idea of where a paper might go, so you can use, for example, either the *ACL template or the ACM template. If you've designed your experiments correctly, this prose will comprise the core of your paper, and your proposed contributions can be captured by additional rows in your results table. This is also consistent with my suggested writing order of a paper.
A few more advantages of this approach: Forcing yourself to accurately describe your baseline often pokes holes in your thinking. Writing is something that you can do while you're waiting for the neural networks to train!
[1] Most? All?
[2] Technically, latex is also code.