Fair evaluations and comparisons are of fundamental importance to the NLP community to properly track progress, especially within the current deep learning revolution, with new state-of-the-art results reported in ever shorter intervals. This concerns the creation of benchmark datasets that cover typical use cases and blind spots of existing systems, the designing of metrics for evaluating the performance of NLP systems on different dimensions, and the reporting of evaluation results in an unbiased manner.
Although certain aspects of NLP evaluation and comparison have been addressed in previous workshops (e.g., Metrics Tasks at WMT, NeuralGen, NLG-Evaluation, and New Frontiers in Summarization), we believe that new insights and methodology, particularly in the last 1-2 years, have led to much renewed interest in the workshop topic. The first workshop in the series, Eval4NLP’20 (collocated with EMNLP’20), was the first workshop to take a broad and unifying perspective on the subject matter. We believe the second workshop will continue the tradition and become a reputed platform for presenting and discussing latest advances in NLP evaluation methods and resources.
Particular topics of interest of the workshop include (but not limited to):
Designing evaluation metrics
Proposing and/or analyzing:
- Metrics with desirable properties, e.g., high correlations with human judgments, strong in distinguishing high-quality outputs from mediocre and low-quality outputs, robust across lengths of input and output sequences, efficient to run, etc.;
- Reference-free evaluation metrics, which only require source text(s) and system predictions;
- Cross-domain metrics, which can reliably and robustly measure the quality of system outputs from heterogeneous modalities (e.g., image and speech), different genres (e.g., newspapers, Wikipedia articles and scientific papers) and different languages;
- Cost-effective methods for eliciting high-quality manual annotations; and
- Methods and metrics for evaluating interpretability and explanations of NLP models
Creating adequate evaluation data
Proposing new datasets or analyzing existing ones by studying their:
- Coverage and diversity, e.g., size of the corpus, covered phenomena, representativeness of samples, distribution of sample types, variability among data sources, eras, and genres; and
- Quality of annotations, e.g., consistency of annotations, inter-rater agreement, and bias check
Reporting correct results
Ensuring and reporting:
- Statistics for the trustworthiness of results, e.g., via appropriate significance tests, and reporting of score distributions rather than single-point estimates, to avoid chance findings;
- Reproducibility of experiments, e.g., quantifying the reproducibility of papers and issuing reproducibility guidelines; and
- Comprehensive and unbiased error analyses and case studies, avoiding cherry-picking and sampling bias.
HumEval invites submissions on all aspects of human evaluation of NLP systems.