How to write paper

Writing, especially top-tier conference, is crucial for acceptance.  You should show excellent logical reasoning skill, show the experts that you know the area well and what is the current frontier of the area, that you possess enough scientific training to execute the research, and that you have already thought deeply and carefully about your topic.

Reviewers do not see when you are working, thus, they can only judge all of these qualities through reading your paper, yes only your few pages of writing, so writing is crucially important.  Even you work so hard on the topic for 1 or 2 whole years or that you have done everything right, but if you cannot explain clearly, reviewers do not care whether you work hard or not.  After all, they can only judge your entire work of one or two years based on your 10+ pages writing.  Here I summarize some key points you need to talk about in your typical paper.

Abstract

The primary function of this section is to describe your work, what did you do and your key finding. For beginners, I recommend this five-part structure with fewer than 200 words.  Using this template can help train your logic in arranging arguments. The first part is about speaking broadly about the importance of your area.  The second part is for describing the problem.  The third part describes what you did.  The fourth part describes your key result.  The last part describes the significance, why I should care.

If you don't believe me, take some papers you read, and check their abstract.  They almost always have these five parts.
 

Introduction

This is the most important section of all.  It's the face. The primary function of this section is to argue your motivation and why did you do this work.  If it sucks, reviewers will likely just skip the rest of the paper and gives you a rejection.  This section usually contains 5 points and 1 key figure  (same structure as abstract but in expanded version).  In writing, the Introduction is just an expansion of the abstract. In the first paragraph, you should explain the background and current state of your area, as well as the core-related work.   In the second paragraph, you should explain the specific problem and gap of past work.  This is perhaps the most important paragraph and you should be very precise what is the problem you want to solve.   In the third paragraph, you should explain what you did.  In the fourth paragraph, you should explain what did you achieve. The last paragraph, it is for explaining the significance and contribution of this work, why we should care, and why it is important for the research community.  For new (especially non-native) students, DON’T DO OTHER FORMAT ASIDE FROM THIS. PLEASE FOLLOW THIS.  It trains you to think logically.  You can try other formats when you get experienced.

Reasons for rejection

  • Not logical, i.e., not supported with evidence.  Each paragraph in the Introduction should be supported with strong evidence.
  • The specified problem is unclear.  The worst problem is about “there is no study who did this, so I want to do it”.  No one did it does not mean it is a good problem.
  • The problem has already been solved or your solution has already been partially suggested by someone.
  • Too common sense or obvious - no impact to the research community
  • The problem is not scoped clearly thus you are over-promising.
     

Related Work

The primary function of this section is 1) to reveal the gap of past work, 2) identify potential hypotheses, and 3) identify experimental design. One good tip is to read some of the papers close to your work and see what they cite commonly; this commonly cited work is usually important to work.  Another good tip is to categorize your related work into different categories, so reviewers can understand it easily.

Reasons for rejection

  • You miss citing a very important work in the field
  • You did not logically reveal the gap or the problem
  • You just list the related work, without giving insights, leaving a lot of "so what" question marks
  • You touch some areas which are not related to your work, thus adding confusion.
  • You try to "teach" too much, but not "synthesize".  Reviewers are experts, they don't want you to teach them.


Models / Devices / Technique / Systems (skip this if your paper is primarily experimental)

The primary function of this section is to describe what you have developed, a software, hardware, an algorithm, etc.  It MUST allow people to replicate your technique.  Usually, you should explain 1) the big goal (where it will be used, and how you want it to be used), 2) the rationale (e.g., why you decide to choose A algorithm rather than B algorithm), 3) any tradeoffs you made, 4) any technical challenges you face, 5) the development process.

Common reasons for rejection

  • You focus too much on implementations but not on the ideas behind the innovation.  
  • You describe inadequately, not allowing people to replicate your work
  • Your solution has little technical novelty nor any special challenges or creativity in it
  • Your solution is not grounded in any principles or rationales.  You did not adequately explain why you think your solution will work, and on what basis you are making that claim.


Method / User Study / Evaluation / Experiment (many titles you can use)

This study section describes the evaluation process.  This is usually the easiest to write.  Simply copy how your role model papers structured and simply put the details.   Most of the paper share the same subsections structure.

Common reasons for rejection

  • Inadequate description - did not fully describe everything such that other people can replicate your experiment
  • Poor experimental design - did not compare what you should have, e.g., baseline, unfair comparisons, the comparisons are not rigorous
  • Inadequate evaluation / metrics - did not really measure what you should have

Results

The primary function of this is to report your results.  Put everything into tables and/or graphs.  Use statistical analysis where appropriate.  

Common reasons for rejection:

  • Inappropriate statistical tests - analyzing your data using the wrong statistical tests or lack of it
  • Bad figures or bad data reporting usually lowers my impression but will not lead to straight rejection.
     

Analysis (optional)

The primary function of section is to analyze the why's of the results.  It could be in the form of ablation study, analyzing the input-output association, analyzing the internal network parameters, etc.   Basically it provides good insights into why the results are like that.

Common reasons for rejection:

  • Inadequate analysis - you simply forget to do some standard analysis that all the people do in your domain.  Otherwise, any extra analysis is always welcomed.

Discussion

The primary function is to discuss why you think the result is like this, did your results match your hypothesis, why sometimes it does not work, what is an important result that I (reviewer) should care and why, do you find anything very surprising? what are the implications of your result?

Common reasons for rejection:

  • Just simply rephrasing the results without any deep insights
  • You propose some guidelines or implications that are too common sense (this is one of the most common mistakes!), i.e., you did not offer any actionable design guidelines or implications
  • You discuss some very big claims without any data support or something that is far away from your data
  • The mismatch between what you promise in the Introduction, and what you discuss.
  • Possible discussion
    • Did you find what you expect?  Discuss in depth anything you did not expect.
    • Did you achieve what you claim?  and what you did not yet achieve? what is your ultimate goal, and how far are you from it?
    • How did your work compare with related work? consistent, anything surprising?
    • Why did you choose certain experimental/approaches design decisions, and how did it affect your results?
    • Design implications or guidelines
    • Significance of your work
    • Limitations: for limitations, please only describe the core limitation that you think might affect the validity of your result. Limitation such as “I conduct this study in the US but I have not tested in the UK” is not a limitation to be discussed as it is almost a common limitation in any study.  Discuss CONCRETE, REAL limitation.
    • Future work, you only have to discuss one or two really concrete one.

Conclusion

The primary function is to 1) confirm that you achieve what you claim in the Introduction, 2) repeat significance of the work, 3) give a strong takeaway message, and 4) a happy ending sentence.

  • Key point is not to discuss any new things; all is simply a summary, please do not add any new terms or ideas in this section
  • Not much impact for rejection or acceptance.

---

Key to success

  • FIND A MODEL PAPER - the easiest way is to find a model paper that is closest to your paper topic and use it as your model.  Learn the style of writing, the contents, and the flow of contents.
  • NICE FIGURES - all good papers have nice-looking, self-explanatory, informative figures.
  • GET FEEDBACK - the only way you can dramatically improve is to get high-level comments from experts. 
  • ITERATE - iteration is the key.  Keep in mind that real writing starts when you revise.
  • WRITE EARLY - the earlier you write, the higher chance you get your work to be accepted.
  • ONLY ONE CONTRIBUTION - remember, a great paper has only ONE contribution.  Having a lot of contributions is usually bad paper, as it has no focus or depth.  Also, putting a lot of contributions in limited space can frustrate readers. A good test is trying to talk your contribution in one sentence such that anyone can understand the value of your work.

Some key writing tips (compiled from common mistakes):

  • Assume readers are very smart.  Write precisely; don't teach, but convince.
  • One paragraph only one idea
  • Choose each word carefully, such that it has no two meanings, e.g., "good" can mean so many things
  • Each paragraph starts with an signal words (“This paper proposes…”) and a signal (e.g., "However") describing the tone and whole idea of the paragraph
  • Each paragraph ends with an ending sentence - this sentence describes the conclusion of this paragraph
  • Avoid non-informative sentence, e.g., "Our work is new."
  • Each sentence should ALWAYS BE BACKED by some evidence, i.e., DON’T MAKE any claims without evidence
  • Make clear what results are surprising, what are expected, do not mix them as it hides away why readers should care about your work. 

Make sure this is clear

  • Are the significance and contribution super clear?
  • Did the work introduce new ideas or approaches?
  • Did you make sure that the results are valid?
  • Did you write very clearly and precisely?
  • Did you read relevant previous work and cite them?

Why LateX

Well, the answer is simple, LaTeX is simply better!  LaTeX is neat, clean and supports many useful libraries.  It automatically figures out the positions of figures/tables, the formatting (e.g., mathematical equation), the inclusion of bib, etc.  LaTeX allows me to focus on what to write, and not how to make it looks good.  You can easily get started with Overleaf.

 

For more readings:

Books I used to learn how to write: