Dataset Card for Multi-LexSum

Talk @ NeurIPS 2022

Dataset Summary

The Multi-LexSum dataset is a collection of 9,280 such legal case summaries. Multi-LexSum is distinct from other datasets in its multiple target summaries, each at a different granularity (ranging from one-sentence “extreme” summaries to multi-paragraph narrations of over five hundred words). It presents a challenging multi-document summarization task given the long length of the source documents, often exceeding two hundred pages per case. Unlike other summarization datasets that are (semi-)automatically curated, Multi-LexSum consists of expert-authored summaries: the experts—lawyers and law students—are trained to follow carefully created guidelines, and their work is reviewed by an additional expert to ensure quality.

Languages

English

Dataset

Data Fields

The dataset contains a list of instances (cases); each instance contains the following data:

Field	Description
id	`(str)` The case ID
sources	`(List[str])` A list of strings for the text extracted from the source documents
summary/long	`(str)` The long (multi-paragraph) summary for this case
summary/short	`(Optional[str])` The short (one-paragraph) summary for this case
summary/tiny	`(Optional[str])` The tiny (one-sentence) summary for this case

Please check the exemplar usage below for loading the data:

from datasets import load_dataset

multi_lexsum = load_dataset("allenai/multi_lexsum", name="v20230518")
# Download multi_lexsum locally and load it as a Dataset object 

example = multi_lexsum["validation"][0] # The first instance of the dev set 
example["sources"] # A list of source document text for the case

for sum_len in ["long", "short", "tiny"]:
    print(example["summary/" + sum_len]) # Summaries of three lengths

  print(example['case_metadata']) # The corresponding metadata for a case in a dict

Data Splits

	Instances	Source Documents (D)	Long Summaries (L)	Short Summaries (S)	Tiny Summaries (T)	Total Summaries
Train (70%)	3,177	28,557	3,177	2,210	1,130	6,517
Test (20%)	908	7,428	908	616	312	1,836
Dev (10%)	454	4,134	454	312	161	927

Dataset Sheet (Datasheet)

Please check our dataset sheet for details regarding dataset creation, source data, annotation, and considerations for the usage.

Additional Information

Dataset Curators

The dataset is created by the collaboration between Civil Rights Litigation Clearinghouse (CRLC, from University of Michigan) and Allen Institute for AI. Multi-LexSum builds on the dataset used and posted by the Clearinghouse to inform the public about civil rights litigation.

Licensing Information

The Multi-LexSum dataset is distributed under the Open Data Commons Attribution License (ODC-By). The case summaries and metadata are licensed under the Creative Commons Attribution License (CC BY-NC), and the source documents are already in the public domain. Commercial users who desire a license for summaries and metadata can contact [email protected], which will allow free use but limit summary re-posting. The corresponding code for downloading and loading the dataset is licensed under the Apache License 2.0.

Citation Information

@article{Shen2022MultiLexSum,
  author    = {Zejiang Shen and
               Kyle Lo and
               Lauren Yu and
               Nathan Dahlberg and
               Margo Schlanger and
               Doug Downey},
  title     = {Multi-LexSum: Real-World Summaries of Civil Rights Lawsuits at Multiple Granularities},
  journal   = {CoRR},
  volume    = {abs/2206.10883},
  year      = {2022},****
  url       = {https://doi.org/10.48550/arXiv.2206.10883},
  doi       = {10.48550/arXiv.2206.10883}
}

Release History

Version	Description
`v20230518`	The v1.1 release including case and source document metadata
`v20220616`	The initial v1.0 release