Paper is an information carrier to present and sell your work，so the paper should be well-written, well-organized, and easily understood. You should make the reviewer happy and buy your contribution.
Overall
- Abstract is the extension of title.
- Introduction is the extension of abstract.
- Conclusion is the rewritten version of abstract.
Language
- Make sure there is no typo and syntax mistake (you can use assistant tools like Grammarly https://app.grammarly.com/ to assist you with this tedious job).
- For technical term, unify them throughout the paper, use at most two sayings throughout the paper. For non-technical term, use at least two sayings alternatively throughout the paper and avoid simple words. Substitutions for common simple words can be found here.
- Your technical writing may look amateur and unprofessional. This problem should be addressed from the following two aspects: domain-invariant writing and domain-specific writing. Even if you have already written dozens of papers in one domain but plan to write a paper in another domain, you also need domain adaptation to fit the new domain.
- Domain-invariant writing: read CV/ML papers in a wide range of fields and learn their paper writing.
- Domain-specific writing: read papers in the related narrow field annd learn their paper writing. Generally, before I write or revise a paper in a new field, I will read at least 10 recent papers and jot down the well-written key words/phrases as raw corpus.
- Merge fragmented short sentences into a coherent long sentence. Use connective words/phrases to glue neighboring sentences to make the transition smoother.
Structure
Writing paper is not randomly putting together what you have in mind. Intuitively, the left subimage is a poorly-written paper and the right one is a well-written paper. Imagine you can enter a cake store and the waitress serves you the left cake, the feeling is the same as that of reviewer when reading a poorly-written paper. The logic relation can be roughly categorized into “and”, “or”, “then”, “but”, “so”, “specifically”.
One paper should be well-structured, composed of hierarchical logic blocks. Writing paper is writing logic flow. Generally speaking, abstract is the extension of title, introduction is the extension of abstract, conclusion is rewritten abstract. In each section, the sentences should be clustered as hierarchical logic blocks in an agglomerative manner.
Content
- Accuracy: Each sentence should exactly present what you want to present and cannot be ambiguous with multiple versions of interpretations.
- Completeness: Your paper should be self-explanatory. You have to ensure that readers do not have to read extra materials to understand your paper. With all necessary ingredients, the remaining problem is the struture issue: how to organize all these ingredients.
- Brevity: Remove redundant expressions and keep it simple. Brevity is the spirit.
Related work
- Classify existing works into several groups. For example, “the existing works can be roughly categorized into XXX, XXX, and XXX.”
- Compared with each group of existing works, state your difference and advantage. For example, “All these methods XXX, but XXX. In contrast, our method XXX”.
Experiment
- For baselines in experimental section, compare with state-of-the-art baselines and yourself (e.g., ablation study or component analysis) qualitatively or quantitatively. It would be great to offer in-depth qualitative analysis.
- When comparing with state-of-the-art baselines, the most important thing is fair comparison. Check whether the experimental settings (e.g., train/test split, features, backbone network) are the same. If the results reported in original papers are from different experimental settings, you need to run their code or even implement their method by yourself.
- If one baseline uses more information and underperforms your method, that is OK.
- If one baseline uses less information and outperforms your method, that is a serious problem.
- If one baseline uses more information and outperforms your method, remove their extra information and compare with it again. Do not leave a better baseline using more information on your paper, which is extremely risky.
- If one baseline uses less information and underperforms your method, see whether this baseline can use the extra information easily. If this baseline can incorporate the extra information in a simple way, compare with it with extra information again.
Math
- Sometimes when it is hard to describe something using natural language, try describing it using math.
- For each variable, please define it before using it. Merge unncessarily duplicated variables and keep as few variables as possible. Think clearly about the meaning of each variable. Sometimes, a variable was defined a long time ago and you need to help reviewer recall that variable.
- Math expressions should be regulated and rigorous.
- For mathematically dense paper, it would be better to start with a general instruction before entering math. For example, “a vector/matrix is denoted by a lowercase/uppercase letter in boldface. The transpose of a vector/matrix is denoted by the superscript ‘. Moreover, $A^{-1}$ is used to denote the inverse matrix of A.
We use $\langle A,B\rangle$ (resp., $A\circ B$) to denote the inner product (resp., element-wise product) of two matrices”.
Show and Hide
- For your contribution, you should highlight it multiple times (e.g., abstract, introduction, related work, method, experiment, and conclusion). Otherwise, the reviewer will claim your contribution is minor.
- For your weakness, try to hide it if possible. If you could not hide it, provide a reasonable and compromising explanation.
Attack and defend
Imagine you are the reviewer, what questions are you going to ask and how the paper could be defended.