Evaluations in machine learning rarely use the latest metrics, datasets, or human evaluation in favor of remaining compatible with prior work. The compatibility, often facilitated through leaderboards, thus leads to outdated but standardized evaluation practices. We pose that the standardization is taking place in the wrong spot. Evaluation infrastructure should enable researchers to use the latest methods and what should be standardized instead is how to incorporate these new evaluation advances. We introduce GEMv2, the new version of the Generation, Evaluation, and Metrics Benchmark which uses a modular infrastructure for dataset, model, and metric developers to benefit from each other’s work. GEMv2 supports 40 documented datasets in 51 languages, ongoing online evaluation for all datasets, and our interactive tools make it easier to add new datasets to the living benchmark.

Data-to-Text Generation (D2T) problems can be considered as a stream of time-stamped events with a text summary being produced for each. The problem becomes more challenging when event summaries contain complex insights derived from multiple records either within an event, or across several events from the event stream. It is important to understand the different types of content present in the summary to help us better define the system requirements so that we can build better systems. In this paper, we propose a novel typology of content types, that we use to classify the contents of event summaries. Using the typology, a profile of a dataset is generated as the distribution of the aggregated content types which captures the specific characteristics of the dataset and gives a measure of the complexity present in the problem. Extensive experimentation on different D2T datasets is performed and these demonstrate that neural systems struggle in generating contents of complex types.