muld / README.md
ghomasHudson's picture
Update README.md (#1)
d2d3b4c
|
raw
history blame
3.75 kB
metadata
annotations_creators:
  - found
  - crowdsourced
language_creators:
  - found
languages:
  - en-US
  - de-DE
  - en
  - de
licenses: []
multilinguality:
  - translation
  - monolingual
pretty_name: The Multitask Long Document Benchmark
size_categories:
  - unknown
source_datasets:
  - original
  - extended|hotpot_qa
  - extended|open_subtitles
task_categories:
  - conditional-text-generation
  - question-answering
task_ids:
  - machine-translation
  - summarization
  - abstractive-qa

MuLD

The Multitask Long Document Benchmark

MuLD (Multitask Long Document Benchmark) is a set of 6 NLP tasks where the inputs consist of at least 10,000 words. The benchmark covers a wide variety of task types including translation, summarization, question answering, and classification. Additionally there is a range of output lengths from a single word classification label all the way up to an output longer than the input text.

Supported Tasks and Leaderboards

The 6 MuLD tasks consist of:

  • NarrativeQA - A question answering dataset requiring an understanding of the plot of books and films.
  • HotpotQA - An expanded version of HotpotQA requiring multihop reasoning between multiple wikipedia pages. This expanded version includes the full Wikipedia pages.
  • OpenSubtitles - A translation dataset based on the OpenSubtitles 2018 dataset. The entire subtitles for each tv show is provided, one subtitle per line in both English and German.
  • VLSP (Very Long Scientific Papers) - An expanded version of the Scientific Papers summarization dataset. Instead of removing very long papers (e.g. thesis), we explicitly include them removing any short papers.
  • AO3 Style Change Detection - Consists of documents formed from the work of multiple Archive of Our Own authors, where the task is to predict the author for each paragraph.
  • Movie Character Types - Predicting whether a named character is the Hero/Villain given a movie script.

Dataset Structure

The data is presented in a text-to-text format where each instance contains a input string, output string and (optionally) json encoded metadata.

{'input: 'Who was wearing the blue shirt? The beginning...', 'output': ['John'], 'metadata': ''}

Data Fields

  • input: a string which has a differing structure per task but is presented in a unified format
  • output: a list of strings where each is a possible answer. Most instances only have a single answer, but some such as narrativeQA and VLSP may have multiple.
  • metadata: Additional metadata which may be helpful for evaluation. In this version, only the OpenSubtitles task contains metadata (for the ContraPro annotations).

Data Splits

Each tasks contains different splits depending what was available in the source datasets:

Task Name Train Validation Test
NarrativeQA ✔️ ✔️ ✔️
HotpotQA ✔️ ✔️
AO3 Style Change Detection ✔️ ✔️ ✔️
Movie Character Types ✔️ ✔️ ✔️
VLSP ✔️
OpenSubtitles ✔️ ✔️

Citation Information

@misc{hudson2022muld,
      title={MuLD: The Multitask Long Document Benchmark}, 
      author={G Thomas Hudson and Noura Al Moubayed},
      year={2022},
      eprint={2202.07362},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Please also cite the papers directly used in this benchmark.