Musings on writing on an NLP/CL paper

Writing an NLP/computational linguistics paper can be a daunting prospect, even for seasoned researchers. I approach it way I approach writing for any academic discipline: the continuing stream of papers in the field can be viewed as one giant colloquy amongst the researchers. The point of that joint conversation is to enhance our collective understanding, by testing hypotheses, uncovering new facts or producing new observations, as well as challenging previous hypotheses or assumptions. Regardless, I try to ask myself, “What am I adding to the conversation?” A useful exercise is this: imagine you are in an actual conversation with researchers whom you admire and who are familiar with the subject area in which you plan to write. What would you want to tell them about your work? Would they get excited about it? How  would you tell them, i.e., what kind of narrative would you create around your work that makes it compelling? Do you think you could convince  them that your work adds something important to the field?

Given that ours is a scientific discipline, once you have been able to answer the above, you must then do your utmost to be diligent. Read as much related work as you can, to become the expert in the area, not only so you can cite all the most relevant work, but also so you can properly contextualize your work as you write it up. Find succinct ways to present your findings, making use of all the best visual representations. If you yourself are not facile in creating visualizations of data, then consult your collaborators or colleagues. Similarly, if you yourself are not the most facile at writing (in English), find people who are, so they can review and help edit your paper. In the final product, make sure it is patently clear what your specific contribution to the colloquy is; some authors go so far as to include a sentence of the form “Our main contribution is…”, and this is not a bad idea.

Finally, I often need to apply a rule of good writing I learned long ago, which is that you may end up writing the beginning of your paper, only to find that the real  first paragraph is hiding two or three paragraphs down. In other words, once you have the first draft of the important introductory paragraphs, consider whether the most direct beginning was buried too far below; promote that to the top by either moving it up or, better still, cutting the extraneous beginning paragraphs entirely. This is similar to the advice on giving a talk that Prof. Stuart Shieber gave in one of the several classes I took from him as an undergrad and graduate student: “A talk is not a murder-mystery,” he said. “You don’t need to keep the audience in suspense until the end. Instead, present your main results at the very beginning.” (I’m paraphrasing slightly.) This is great advice for a scientific paper as well as a talk. We love to tell a good story with a surprising ending; however, while narrative is still very important even in scientific writing, the conclusions come at the beginning, followed by the “story” that led to them.

Thanks for reading, and happy writing!


Dan Bikel (Google Research)