Skip to main content

22. References, Appendices, and Reproducibility

The back matter is where a paper earns trust after it has made its case. Three things live here, each with a different job: references let a reader verify the conversation you joined, appendices hold what supports your claims without being essential to them, and the reproducibility record lets someone rebuild your results. None of the three is a dumping ground — material the main claims depend on belongs in the main text, and material that supports no claim at all belongs nowhere.

The appendix test is one question: can a reviewer accept the paper's claims without turning to the appendix? If yes, the content is placed correctly. If a claim only holds once the reader reaches Appendix C, that content is in the wrong place. This mirrors the boundaries drawn earlier in the guide — the method goes in Methods, hyperparameter values in Experimental Setup, data-collection detail in Dataset Construction — and the appendix takes the overflow from each that is useful but inessential.

References​

  • Alphabetical by first author. ACL convention.
  • Use the ACL Anthology BibTeX wherever the paper appears there. The bib keys are stable and other authors will find your citations easier to verify. ACL Anthology bib key format: {firstauthor}-{etal}-{year}-{titleword}.
  • Cite the peer-reviewed version, not the arXiv preprint, if both exist. ACLPUB guidelines are explicit on this.
  • Do not hallucinate citations. This is now a documented problem in the field — hundreds of recent *ACL papers contain references to nonexistent works, most produced when authors used LLMs to help draft without verifying the output. Verify every citation. If you used an LLM to help draft, check every reference it produced.

Appendices​

Appendices go after references. They do not count toward the page limit, but they should not contain content essential to the main claims — reviewers are not required to read them.

Good appendix content:

  • Hyperparameter tables.
  • Additional ablation results.
  • Dataset construction details (annotation instructions, screenshots of interfaces).
  • Proofs and derivations.
  • Additional qualitative examples.

Bad appendix content:

  • Your main results table (it belongs in the main text).
  • The definition of your method (also belongs in the main text).
  • An eight-page literature review you could not fit elsewhere.

Reproducibility​

ACL has a formal Responsible NLP Research Checklist that you must fill out at submission. Read it now, not the night before the deadline: https://aclrollingreview.org/cfp.

Key reproducibility items:

  • Code and data release. Anonymized for the review period; deanonymized for camera-ready. Use an anonymous GitHub mirror or anonymous.4open.science during review.
  • Random seeds and number of runs.
  • Compute used. Number and type of GPUs, total training time.
  • Hyperparameter search. Range searched, criterion, best value.
  • Licenses. For all data, models, and code used.

A practical workflow​

About one week before the deadline, freeze your code and your experiments. Tag the commit. From that tag, you should be able to reproduce every number in the paper with one command. If you cannot, write the documentation to make it possible. Reviewers may not look — but they may, and if they do, this is the difference between a clean accept and a "results may not be reproducible" rejection.

Further reading​