BEGIN

https://github.com/google/BEGIN-dataset

A benchmark dataset for evaluating dialog system and natural language generation metrics. See our paper "Evaluating Groundedness in Dialogue Systems: The BEGIN Benchmark" at https://arxiv.org/abs/2105.00071 .

BEGIN is a Benchmark for evaluating response attribution to some given background knowledge. It is comprised of 12k dialogue turns generated by state-of-the-art dialogue systems trained on three widely used knowledge-grounded dialogue benchmarks, and annotated by humans for the attribution of responses.

Using BEGIN, we analyze the robustness of 8 evaluation metrics. We find that simply resorting to correlation with human judgment is not sufficient to draw conclusions about the effectiveness of the metrics. More specifically, we observe that these metrics rely heavily on spurious correlations and perform poorly on attributable abstractive responses confounding them with unattributable ones. Finally, we find that these metrics perform substantially worse when tested on benchmarks with long knowledge sources, corroborating inherent robustness issues under distribution shift.

This is not an officially supported Google product. This means that Google may not regularly release updates or provide support in conjunction with your use of this dataset.

Authors

• Nouha Dziri
• Hannah Rashkin [email protected]
• Tal Linzen [email protected]
• David Reitter [email protected]

Data Format

We split BEGIN into 90% test set and 10% development set, and store them in a tsv (tab-separated) format. Below, we define each column:

model name          The dialogue model used to generate responses
                    (gpt2, t5, ctrl [CTRL-Dialog], doha)
data_source         The dialogue benchmark used to train the models
                    (wow [Wiz. of Wikipedia], cmu [CMU DoG], tc [Topical Chat])
knowledge           The knowledge-snippet on which the response should be grounded
message             The previous turn in the dialogue history
response            The generated response
begin_label         BEGIN label (Fully attributable, Not fully attributable, Generic)

Documentation

Note that the dataset has been revised as compared to the initial preprint.

https://arxiv.org/abs/2105.00071

Citation

This work has been submitted to TACL, and we are still awaiting the decision. In the meantime, please cite our work if you use BEGIN in your research.

@article{dziri2021evaluating,
  title={Evaluating groundedness in dialogue systems: The begin benchmark},
  author={Dziri, Nouha and Rashkin, Hannah and Linzen, Tal and Reitter, David},
  journal={arXiv preprint arXiv:2105.00071},
  year={2021}
}

Versions

• Example release 6/1/2021: Only examples included
• First full release 6/1/2021
• Second full release 6/1/2022: New data collection methodology
  • We changed to a simplified annotation question design.
  • We added output from two additional models (Doha and Ctrl-Dialog).
  • We added two datasets, TopicalChat and CMU-Dog.