(Work in progress, send a PR!)
- When in doubt use sections -> Introduction, Background, Model, Training, Methods, Results, Discussion, Conclusion.
- Tables should always follow this guide
- Tables / Figures should always float. Never inline in the text.
- When using natbib, \citet is for when the citation is a noun, and \citep is for when it is at the end.
- Captions should be short but fully self-explanatory of the columns / rows. They should not use 1st person.
- Abstracts should be 1 paragraph. When in doubt -> Context, Problem, Idea 1, Idea 2, Results.
- Section titles should be starting-caps.
- The goal of related work is not just to list papers, but to explicitly make claims as to how your work differs from each one.
- Figures should have a white background and large fonts. Do not screenshot! Generate a high-res, pdf output.
- Use present tense (almost) everywhere.
- You do not need a summary paragraph at the end of your intro.
- All empirical results must be in a table or figure.
- Methods section should not introduce new modeling. Enumerate the tasks, baselines, hyperparameters.
- Results section should not introduce new tasks or models. Summarize the tables.
- Any non-trivial notation should be introduced as early possible. Ideally background.
- 8 pages is an extremely hard limit.
- Always use `` '' for quotes not " ".
- Use bold sparingly. Opt for italics for new technical terms.
- Turn off
\usepackage[review]{emnlp}to\usepackage[]{emnlp}while editing to fix overleaf linking. - Use
\newcommand{\todo}[1]{{\small\color{red}{\bf [*** Todo: #1]}}}for inline comments.
- Rational Reconstruction https://web.stanford.edu/class/cs224u/readings/shieber-writing.pdf
- ICML paper blog https://icml.cc/Conferences/2002/craft.html
- What are the 3 contributions of the paper?
- Do my experiments convincingly prove each of these are true?
- Can I cut anything that does not satisfy these?
- Would someone who has not read a paper in 2 years understand what is happening?