Guest Post: Reproducibility at EMNLP 2020
This is a guest post by Jesse Dodge and Noah A. Smith explaining the rationale behind the Reproducibility checklist at EMNLP 2020 (shown in our call for papers), and a plan for a reproducibility challenge after the conference.
– EMNLP PC Chairs
In 2020, natural language processing is largely an experimental field, and the validity of our field's conclusions rests on the reproducibility of our experiments. NLP researchers hold a range of different positions about which questions are the most pressing and which applications have the greatest potential to aid humanity. Yet, as a scientific community we have a shared interest in the generalizability of published results. There are many factors that influence how reproducible our conclusions are and how well our experimental findings generalize, but these factors are often underreported when our published papers focus primarily on presenting new results. As a premier publication forum in our field, the EMNLP conference is in a position to encourage and incentivize improved reporting of the information necessary to reproduce published experimental research. We start from the position that a submission to EMNLP should contain sufficient information to allow future researchers to independently reach the same conclusions (for example, about model performance).
Our focus is on improved reporting of the setup and results of the experiments that authors have conducted, so that a paper will detail the process by which the authors came to their conclusions and allow future researchers to understand and build on their work.
We introduce a reproducibility checklist for NLP (shown in the EMNLP 2020 call for papers). Our checklist builds on the machine learning reproducibility checklist, but is refocused for NLP papers. The machine learning reproducibility checklist that will be used at NeurIPS 2020 has aligned some items with ours; we plan to quantitatively analyze our checklist responses, and this cross-referencing will allow us to compare across communities.
The EMNLP 2020 program committee does not require that any items on the checklist be included in the paper, only that the checklist be filled out by authors. The filled-out checklist will not be released with the published version of an accepted paper, it is meant as a tool for authors and reviewers. There is great variety in the research papers submitted to our conferences, and some items on the checklist will be appropriate for only some papers, so "N/A" is an option for all responses. For example, authors of a submission that introduces a new dataset but doesn't tune any hyperparameters would respond N/A to the hyperparameter tuning items on the checklist.
A main goal of the checklist is to remind authors about scientifically relevant items to include in their papers; there is a lot to report, and it's easy to forget some details when we're focused on explaining and writing up new results.
Some items on the checklist might be surprising. For example, why report development set performance? Consider a practitioner trying to reimplement your approach. If the only performance scores reported are from the test set, then they will likely need to repeatedly evaluate models on test data just to verify their implementation. However, test sets should be consulted minimally as an estimate of generalization performance, so development set performance is a preferable tool for this use case.
The checklist also includes the average runtime for each approach, and the number of parameters in each model. Runtime is a notoriously fickle measure, but coupled with a description of the computing infrastructure used (another item on the checklist), it can provide useful information for a reader who wants an estimate of the computational resources required to use the methods in the paper. All together, these items are a step toward reporting time and space complexity of our algorithms, which are foundational ideas in computer science.
A typical NLP pipeline involves evaluating many models (varying the model architecture, hyperparameters, etc.) on the development set, and reporting the best-found performance. Only reporting this single point discards the results of all-but-one of the finished experiments. These other trials aren't necessarily negative results; instead they reflect the process which led to the publishable findings. Because there is so much variety in the methods and budgets used in our field, readers are often in the dark. The checklist items for hyperparameter tuning results are meant to illuminate the authors' process.
One item on the checklist is to include source code. This is perhaps the most straightforward path for others to build upon published work, and to replicate the specific experiments in a paper. Public code has facilitated much progress in our field, and as a guide for what to include in a codebase, Papers With Code recently released a set of data-driven recommendations which can be found here.
The main focus of NLP papers is to present new research, and our submissions have limited length. Some authors may find the appendix an appropriate place to include some of these important technical details which underlie our work.
Reproducibility in NLP is a multi-faceted challenge that this first checklist only begins to address. Improving methodological rigor of scientific work is, like science itself, an incremental process. As a next step after EMNLP 2020, we plan to run the first NLP Reproducibility Challenge, an open activity aiming to empirically evaluate the reproducibility of findings presented at EMNLP, and to identify remaining impediments to reproducibility.
Authors of a "reproduction" will choose a single accepted paper from EMNLP 2020 and attempt to reproduce the experiments supporting the main claims from that paper. The authors will report on whether the attempt was successful, and on what obstacles they found. They will be encouraged to enumerate the necessary details for replicating the original work, such as hyperparameter values and the number of GPU hours required to rerun the experiments. Reproductions submitted as a part of the challenge will be made publicly available.
The collection of reproductions will act as a useful public resource; readers of a published paper will be able to read the associated reproduction report to find relevant information which may have not made it into the original paper. The focus of these reproductions will be on clear and appropriate experimental design; with these goals met, succeeding or failing to reproduce the results from the original paper provides valuable information.
In addition to providing the reproductions as a resource to the community, writing a reproduction is an excellent educational opportunity for students to walk through the research process guided by a published paper. We ran a small-scale reproducibility challenge in the Winter 2020 graduate-level NLP course at the University of Washington, in which teams reproduced results from papers published at EMNLP 2019. Students were provided with a detailed, structured template to help ensure that their reports would include clear information about experimental design, computational budget, and results found from their experiments.
Most of the reproductions were successful, though reproducibility in NLP can be challenging, and the students often had to experiment before they were able to reproduce the original results. This provided an opportunity for the students to highlight sources of variation they found, such as reporting training curves or architectural choices, and to examine whether the conclusions of the original paper held up with these variations. This exercise directly informed our design of the EMNLP 2020 checklist and the 2020 NLP Reproducibility Challenge.
More details about the Reproducibility Challenge will be announced later.
Over the past few decades, our field has blossomed from a small, tight-knit community where methodological best practices were essentially an oral tradition passed from advisor to student, student to student, and reviewer to author. This made it easy for new ideas to be rapidly considered, explored, and adopted. Today, the field is both increasingly distributed and connected to other fields. Some community-wide norms exist, but NLP researchers around the world have diverse perspectives on the factors leading to sound empirical research. As our field grows and evolves, so too must we update our norms. Establishing methodology as a central part of our scholarly discourse will enable informed decisions and consistent community expectations. This will not only improve the quality of our work, but also make it easier for students and other newcomers to the field to begin to meaningfully contribute. We see an evolving checklist as a useful tool (much like review forms and calls for papers) to help structure that discourse.