Datasets:

NightRaven
/

bangla-emergency-posts

language:
  - bn
license: cc-by-4.0
task_categories:
  - text-classification
task_ids:
  - multi-class-classification
pretty_name: Bangla Emergency Posts
size_categories:
  - 1K<n<10K
tags:
  - bangla
  - bengali
  - emergency
  - social-media
  - crisis-informatics
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.csv
      - split: validation
        path: data/validation.csv
      - split: test
        path: data/test.csv
  - config_name: full
    data_files:
      - split: full
        path: data/bangla_emergency_posts.csv

Bangla Emergency Posts

5,836 Bangla social media posts, hand-labelled into nine emergency categories. Built for Bangla Emergency Post Classification on Social Media using Transformer Based BERT Models (EICT 2023).

Emergency text classification in Bangla is scarce despite the language having hundreds of millions of speakers. This dataset exists so that Bangla-speaking people can report emergencies in their own language and have those reports routed automatically.

Usage

from datasets import load_dataset

ds = load_dataset("NightRaven/bangla-emergency-posts")
print(ds["train"][0])
# {'content': 'ভূমিকম্পের পর স্কুলবাড়ি ভেঙে পড়ে। ...', 'label': 'natural_disaster'}

Fields

field	type	description
`content`	string	the post text, Bangla with occasional English words
`label`	string	one of the nine categories below

Labels are stored as strings. Where an integer id is needed, the models trained on this data use alphabetical order:

id	label	id	label	id	label
0	accident	3	fire	6	suicide
1	blood	4	natural_disaster	7	war
2	crime	5	pandemic	8	weather

blood covers urgent blood-donation appeals, a common and distinct category of emergency post in Bangla social media.

Splits

Stratified 56 / 14 / 30.

label	train	validation	test	total
accident	579	123	301	1,003
blood	123	31	66	220
crime	1,355	372	765	2,492
fire	295	82	122	499
natural_disaster	277	64	153	494
pandemic	83	24	40	147
suicide	114	33	61	208
war	104	20	51	175
weather	337	70	191	598
total	3,267	819	1,750	5,836

Collection and labelling

Posts were gathered from Facebook, Twitter and Bangladeshi daily newspapers, covering events in Bangladesh and neighbouring India. Labelling was done by hand by native Bangla speakers.

Collecting structured Bangla social media text is difficult — stemming rules are hard to pin down and stopword lists are inconsistent — so the data was labelled manually rather than filtered heuristically.

Citation

@inproceedings{nabil2023bangla,
  title     = {Bangla Emergency Post Classification on Social Media using
               Transformer Based BERT Models},
  author    = {Nabil, Alvi Ahmmed and Arifeen, Shamsul and Das, Dola and
               Salim, Md. Shahidul and Fattah, H. M. Abdul},
  booktitle = {6th International Conference on Electrical Information and
               Communication Technology (EICT)},
  address   = {Khulna, Bangladesh},
  year      = {2023}
}

Models trained on this dataset: NightRaven/bangla-emergency-post-classification