What is the paper for?

The system description paper should enable other researchers to:

• Verify what the system does and how it has been trained.

• Reimplement the system to reproduce the results.

• Understand the system’s strengths and weaknesses.

How should the paper be structured?

A common structure for system description papers is as follows:

• Abstract: four/five sentences highlighting your approach and key results.

• Introduction: ¾ a page expanding on the abstract mentioning key background such as why the task is challenging for current modeling techniques and why your approach is interesting/novel.

• Data: review of the data you used to train your system. Be sure to mention the size of the training, development and test sets that you’ve used, and the label distributions, as well as any tools you used for preprocessing data.

• System: a detailed description of how the systems were built and trained. If you’re using a neural network, were there any pre-trained components, how was the model trained, what hyperparameters were chosen and experimented with? How long did the model take to train, and on what infrastructure? Linking to source code is valuable here as well, but the description should be able to stand alone as a full description of how to reimplement the system. While other paper styles include background as a separate section, it’s fine to simply include citations to similar systems which inspired your work as you describe your system.

• Results: a description of the key results of the paper. If you have done extra error analysis into what types of errors the system makes, this is extremely valuable for the reader. Unofficial results from after the submission deadline can be very useful as well.

• Discussion: general discussion of the task and your system. Description of characteristic errors and their frequency over a sample of development data. What would you do if you had another 3 months to work on it?

• Conclusion: a restatement of the introduction, highlighting what was learned about the task and how to model it.

Developing and running a shared task is a lot of effort. Please acknowledge the contributions of the organizers by citing the following papers:

• For the shared task description paper, which is the main reference for the new test set, use:

@inproceedings{alhafni-etal-2022-gender-rewriting,

title = "The Gender Rewriting Shared Task",

author = "Alhafni, Bashar and Obeid, Ossama and Bouamor, Houda and Habash, Nizar",

booktitle = "Proceedings of the Seventh Arabic Natural Language Processing Workshop",

year = "2022",

address = "Abu Dhabi, United Arab Emirates"

}

• For the Arabic Parallel Gender Corpus for used for training and development, use:

@inproceedings{alhafni-habash-bouamor:2022:LREC,

title = "The Arabic Parallel Gender Corpus 2.0: Extensions and Analyses",

author = "Alhafni, Bashar and Habash, Nizar and Bouamor, Houda",

booktitle = "Proceedings of the Language Resources and Evaluation Conference",

year = "2022",

address = "Marseille, France",

publisher = "European Language Resources Association",

pages = "1870--1884",

}

Acknowledgment

This page is an adaptation of the guidelines shared by the organizers of the First Workshop on Gender Bias for Natural Language Processing at ACL 2019.