Git Product home page Git Product logo

begin-dataset's Introduction

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

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.

begin-dataset's People

Contributors

davidswelt avatar

Stargazers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar

Watchers

 avatar  avatar  avatar

begin-dataset's Issues

wow error (graceland)

I think this isn't fully attributable because there is no way to decide that "it" in the evidence refers to Graceland, but the answer does resolve to it.

it is located at 3764 elvis presley boulevard in the vast whitehaven community, about 9 miles (14.5 km) from downtown and less than four miles (6 km) north of the mississippi border.

oh isn't graceland in memphis? i want to visit some time.

it is located at the elvis presley boulevard

Fully attributable

error in wow data

This one isn't fully attributable (because of equivalent... "Yes it is"):

doha wow It was created by Joy Whitby. Is that like the equivalent of Barney in the US? Yes it is , it was created by Joy Whitby . Fully attributable

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.