NightRaven/bangla-emergency-post-classification

Bangla Emergency Post Classification

Models that classify Bangla social media posts into nine emergency categories, so that posts needing attention from emergency services, local authorities or law enforcement can be surfaced automatically.

Trained for Bangla Emergency Post Classification on Social Media using Transformer Based BERT Models, 6th International Conference on Electrical Information and Communication Technology (EICT), Khulna, Bangladesh, December 2023.

Labels

Nine classes, encoded in 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, which are a common and distinct category of emergency post in Bangla social media.

Data

5,836 Bangla posts collected from Facebook, Twitter and daily newspapers, labelled by hand by native speakers. Posts frequently mix English words into Bangla script.

split	posts
train	3,267
validation	819
test	1,750

The classes are heavily imbalanced — crime is 42.7% of the data and pandemic 2.5%:

label	count		label	count
crime	2,492		suicide	208
accident	1,003		war	175
weather	598		pandemic	147
fire	499
natural_disaster	494		total	5,836
blood	220

What is in this repository

Transformer/
  BanglaBERT/social-media_weights.pt     sagorsarker/bangla-bert-base
  XLM-RoBERTa/social-media_weights.pt    xlm-roberta-base
  mBERT/social-media_weights.pt          bert-base-multilingual-uncased
dnn/
  BiLSTM.model/  BiLSTM_CNN.model/  LSTM.model/    TensorFlow SavedModel
machine_learning/
  logistic/  knn/  mnb/                            scikit-learn, joblib

Usage

The transformer checkpoints are state_dicts for the wrapper classes in the project's modelss/ package, not plain Hugging Face models — loading them with AutoModelForSequenceClassification.from_pretrained will not work. Use the code from the project repository:

from huggingface_hub import hf_hub_download
import torch
from modelss.text_xlm_roberta import TextRoBERTa

model = TextRoBERTa(pretrained_model='xlm-roberta-base', num_class=9, fine_tune=True)
weights = hf_hub_download(
    repo_id="NightRaven/bangla-emergency-post-classification",
    filename="Transformer/XLM-RoBERTa/social-media_weights.pt",
)
model.load_state_dict(torch.load(weights, map_location='cpu'))
model.eval()

Inputs are tokenised at a maximum sequence length of 100.

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}
}