A Comparative Analysis of Noise Reduction Methods in Sentiment Analysis on Noisy Bangla Texts

This repository contains the code and datasets used in the paper titled "A Comparative Analysis of Noise Reduction Methods in Sentiment Analysis on Noisy Bangla Texts " accepted in The 9th Workshop on Noisy and User-generated Text (W-NUT) collocated with EACL 2024.

Paper Link: https://www.arxiv.org/abs/2401.14360.

Table of Contents

• Models
• Datasets
• Training & Evaluation
• Benchmarks
• Future Research Directions
• License
• Citation

Models

The code used to train and test the models are available in this repository under the folder NC-SentNoB Codes.

Datasets

The NC-SentNoB dataset is available at this repository on the folder NC-SentNoB Dataset.
Also available on: [HuggingFace] | [Paperswithcode] | [Kaggle]

SentNoB dataset with Back translation applied is also available on the folder Back-Translated Data. The 1000 ground truths used to evaluate denoising methods are available on the file 1000 Ground Truth.xlsx

Training & Evaluation

We used seven pretrained transformer models for Sentiment Analysis and SVM, BiLSTM, BanglaBERT for Noise Identification.

Benchmarks

• Noise Identification

Model	Precision	Recall	F1-Score
SVM (C)	0.76	0.45	0.57
SVM (W)	0.64	0.38	0.48
SVM (C+W)	0.75	0.45	0.56
Bi-LSTM	0.36	0.18	0.24
Bangla-BERT-base	0.73	0.54	0.62

• Sentiment Analysis on Noisy Text

Model	Precision	Recall	F1-Score
Bangla-BERT-Base	0.72	0.72	0.72
BanglaBERT	0.75	0.75	0.75
BanglaBERT Large	0.74	0.74	0.74
BanglaBERT Generator	0.72	0.72	0.72
sahajBERT	0.72	0.72	0.72
Bangla-Electra	0.68	0.68	0.68
MuRIL	0.73	0.73	0.73

• Sentiment Analysis after Noise Reduction

Model	Precision	Recall	F1-Score
Bangla-BERT-Base	0.69	0.69	0.69
BanglaBERT	0.72	0.72	0.72
BanglaBERT Large	0.73	0.73	0.73
BanglaBERT Generator	0.70	0.70	0.70
sahajBERT	0.70	0.70	0.70
Bangla-Electra	0.66	0.66	0.66
MuRIL	0.71	0.71	0.71

Future Research Directions

1. Develop robust noise reduction models for Bangla texts.
2. Investigate and develop noise specific reduction techniques.
3. Leverage LLMs for noise reduction.
4. Compare performance of Sentiment Analysis using LLMs in both settings: with and without noise.
5. Implement character-level NMT models for back-translation to reduce noise.

License

Contents of this repository are restricted to non-commercial research purposes only under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

Citation

If you use any of the datasets, models or code modules, please cite the following paper:

@misc{elahi2024comparative,
      title={A Comparative Analysis of Noise Reduction Methods in Sentiment Analysis on Noisy Bangla Texts}, 
      author={Kazi Toufique Elahi and Tasnuva Binte Rahman and Shakil Shahriar and Samir Sarker and Md. Tanvir Rouf Shawon and G. M. Shahariar},
      year={2024},
      eprint={2401.14360},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}