How to write a quilified academic paper?

This blog is a water-tight version of Zhiyuanliu's tutorial on how to write a qualified NLP paper.

Process of Paper Publication

Usually, a classic process of paper publication is :

Proposal \(\longrightarrow\) Model Design \(\longrightarrow\) Coding \(\longleftrightarrow\) Parameter Tuning \(\longrightarrow\) Paper Writing \(\longleftrightarrow\) Paper Reediting \(\longrightarrow\) Paper Sumbmitting \(\longrightarrow\) Presentation

An excellet paper = An excellent work (step 1) + An excellent Writting (step 2)

Classic Framwork of a NLP paper

Most of the conference papers (8-10 pages) should include 6 components: Abstract, Introduction, Related Work, Method, Experiment, and Conclusion.

• Abstract: Using 100-200 words to describe clearly the task, existing challenges, your methodology, your experiment results, and finally, your conclusion.
• Introduction: Leaving 1 page to detail the research problem, existing state-of-the-arts, main challenges, typology of methods, specific methods in each branch, and overall performance of each method.
• Related Work: Using 0.5-1 page for related works, distinguish the differences between your work and others.
• Method: Using 2-3 pages to introduce all details of your proposal.
• Experiment: Using 2-3 pages to recommend the datasets, experimental criteria, parameter setting, experimental results, and analysis of these results.
• Conclusion: Summing up the main work and contributions, and proposing some possible directions for the future.

tips:

1. Writing in the interest of readers. Remember to describe all new concepts in your paper, and detail the work to the extend that can your reader easily to grap your idea.
2. Writing with rigorous logic. Guarantee the consistency and coherence of your chapeters, paragraphs, and sentences.

For example. (1) If you mentioned 3 challenges in your introduction, then, make sure you have 3 components in your proposal to deal with all challenges, 3 algorithms in your method part, and 3 types of experiments to verifying each of your algorithms; (2) You should clear the strctures of your paragraphs, whehter they are organized parallelly, or incrementally. Make sure there is a sentence or adverb to connect paragraphs or sentences.

How to write Abstract and Introduction part

Introduction

1. Problem description and the importance of solving it. For a well-known problem, you don't need to emphasis much. However, for a new research problem, you should write more words to highlight the importance of this problem.
2. Existing Works. Standing upon the gaint's shoulder, introduce the representative of all methods. You are going to improve the representative.
3. Presenting challenges and shortcomings of the representative. Make objective criticism on those works.
4. How can your method solve those challenges you mentioned before?
5. Proofing that your method can work better considering those challenges.
6. Summing up the main contributions or your work with precise and concise sentences.

Abstract

Abstract can be view as a brief of your introduction. But you should paraphrase the content in introduction part. You can see clearly the conrespondence between abstract and introduction.

How to write Method part

1. You should introduce the total framework of your method, which offers an overview of your work. Then, give a mathematical formula to describe your prolbem if possible. All notations you going to use should be listed. Finally, introduce all components of your method.
2. You detail every component from the whole to the part, and bold the title of each part. This will navigate the reader to look through your content with a clear clue.

Make the description of framework different from what has been mentioned in introduction part. The description here should be correlated with that in introduction part, but offers more details. Another thing is that you should explain every new notation explicitly.

How to write Experiment part

Before posting the experimental results, you should mention the datasets you used, those criteria you chosen, and all baselines you compared in your expriments.

For presenting your experiments. You should mainly focus on valid the effectiveness of your method on the challenges you dealing with. That's means your method should be proofed to improve the performance after using it. However, Comparison with the most recent state-of-the-art is not required.

In order to present the advantage and characteristic of your method. You can do some experiment to test the hyperparameter effects, Ablation Test (how much contribution each component made), Impact of Datsets (few-shot, class-imbalanced), Error Analysis (limitation of your method, where your method fails), Case study (strength of your method, where your method works best) and so on.

How to write Related Work part

The function of this part is to conclude, classify, and analyse the existing works. The content can be organized chronically or taxonomically. The content should emphasis the challenges they tackled, and the problems they didn't fix. Finally, you should highlight the differences of your method compared with others.

How to write Conclusion

Conclusion is used to summarize contributions and the results to support those contributions. Additionally, one may further possible research topics for future.

参考 (References)

1. https://mp.weixin.qq.com/s/X48Cm58eub_sULlLRk-VIA