Green and Sustainable NLP
This EACL edition is proud to introduce a new track to the main session: Green and Sustainable NLP. We are excited to welcome the Green and Sustainable NLP PC chairs: Roy Schwartz and Emma Strubell, who will manage this unique track and who have kindly contributed some introductory words for the track:
The amount of computation required for deep learning research in NLP has increased dramatically in the past few years. This has led to a surprisingly large carbon footprint for NLP experiments, and to the financial cost of running experiments being increasingly burdensome for academics, students, and researchers, particularly those in emerging economies. These trends are the direct result of the NLP community primarily valuing state-of-the-art results, as opposed to comparing methods with varying amounts of data, model parameters, and hyperparameter trials. The Green NLP track at EACL 2021 aims to promote efficiency as a core evaluation criterion for NLP models alongside accuracy and related measures, to improve community norms around reporting experimental results (e.g. reporting training curves for large, pretrained models), and to facilitate more robust comparisons between approaches with varying amounts of resources.
Specifically, our goal is to encourage submissions relating to:
- Improved reporting of experimental results as we vary the amount of available computation, such as the size of the model, the amount of training data, or the number of hyperparameter optimization trails.
- Methods which present efficient solutions to NLP problems. These include, among others, models that are efficient in: Training data, Training time, Inference time, Number of hyperparameter trails.
- Evaluation criteria which promote efficiency and improve comparisons between models with varying computational budgets. Many different criteria for efficiency exist, such as runtime, number of parameters, floating point operations, and dollar cost. It’s often unclear, however, which criteria to use, and some criteria are difficult to measure. We hope to make progress by encouraging the development of easy-to-use solutions for measuring efficiency.