README.md

metadata

language:
  - en
  - ja
  - ko
  - si
  - ta
pretty_name: M2DS
tags:
  - multilingual summarisation
  - multi-document summarisation
  - dataset
  - nlp
  - bbc
task_categories:
  - summarization
size_categories:
  - 10K<n<100K
configs:
  - config_name: english
    default: true
    data_files:
      - split: train
        path: english/train.json
      - split: validation
        path: english/validation.json
      - split: test
        path: english/test.json
  - config_name: japanese
    data_files:
      - split: train
        path: japanese/train.json
      - split: validation
        path: japanese/validation.json
      - split: test
        path: japanese/test.json
  - config_name: korean
    data_files:
      - split: train
        path: korean/train.json
      - split: validation
        path: korean/validation.json
      - split: test
        path: korean/test.json
  - config_name: sinhala
    data_files:
      - split: train
        path: sinhala/train.json
      - split: validation
        path: sinhala/validation.json
      - split: test
        path: sinhala/test.json
  - config_name: tamil
    data_files:
      - split: train
        path: tamil/train.json
      - split: validation
        path: tamil/validation.json
      - split: test
        path: tamil/test.json

M2DS v1.0 — Multilingual Dataset for Multi-document Summarisation

M2DS is a multilingual multi-document summarisation dataset built from BBC news articles and professionally written BBC summaries across five languages: English, Japanese, Korean, Sinhala, and Tamil.

Quick start

from datasets import load_dataset

# Load a specific language
ds = load_dataset("KushanH/m2ds", "english")

# Access splits
train = ds["train"]
val   = ds["validation"]
test  = ds["test"]

# Inspect a single example
print(train[0]["document"])  # concatenated source articles
print(train[0]["summary"])   # reference summary

Available config names: english, japanese, korean, sinhala, tamil.

Dataset structure

Each language is released as split-based files compatible with Hugging Face load_dataset().

Splits

Split	Purpose
train	Model training
validation	Hyperparameter tuning
test	Final evaluation

Fields

Each row represents one multi-document cluster and contains two fields:

Field	Type	Description
document	string	Multiple related source articles concatenated into one text field
summary	string	Reference summary combining BBC summaries for the cluster

Document separator

Within the document field, individual articles are separated by:

|||||

Example:

Article one text here... ||||| Article two text here... ||||| Article three text here...

Split ratios

• English: 80 / 10 / 10
• Japanese, Korean, Sinhala, Tamil: 90 / 5 / 5

Statistics

Language	Train	Validation	Test	Total	Paper
English	13,496	1,688	1,687	16,871	17K
Japanese	9,891	549	551	10,991	11K
Korean	7,021	391	390	7,802	8K
Sinhala	4,942	275	275	5,492	5.5K
Tamil	8,916	495	496	9,907	10K
Total	44,266	3,398	3,399	51,063	~51.5K

Paper-reported values are rounded per-language presentation values.

External resources

• OSF Archive: https://osf.io/7gjtm/
• GitHub Repository: https://github.com/KushanMH/m2ds

Citation

If you use M2DS in your research, please cite:

@inproceedings{hewapathirana2024m2ds,
  title={M2DS: Multilingual Dataset for Multi-document Summarisation},
  author={Hewapathirana, Kushan and de Silva, Nisansa and Athuraliya, CD},
  booktitle={International Conference on Computational Collective Intelligence},
  pages={219--231},
  year={2024},
  organization={Springer}
}