rasyosef/amharic-news-category-classification

Amharic News Category Classification

This amharic text dataset can be used to train/finetune models for the following tasks

• classification : using the categories
• summarization : using the headlines

Finetuning

Here is a github repo that contains three notebooks that use this dataset to finetune the following models.

• xlm-roberta-base : a multilingual transformer model with 280M parameters
• bert-small-amharic : a new amharic version of the bert-small transformer model with 25.7M parameters, pretrained from scratch using unlabelled amharic text data
• bert-mini-amharic : a new amharic version of the bert-mini transformer model with 9.67M parameters, pretrained from scratch using unlabelled amharic text data

https://github.com/rasyosef/amharic-news-category-classification

The finetuned model classifies a given Amharic news article into one of the following 6 categories.

• ሀገር አቀፍ ዜና (Local News)
• መዝናኛ (Entertainment)
• ስፖርት (Sports)
• ቢዝነስ (Business)
• ዓለም አቀፍ ዜና (International News)
• ፖለቲካ (Politics)

Fine-tuned Model Performance

Since this is a multi-class classification task, the reported precision, recall, and f1 metrics are macro averages.

Model	Size (# params)	Accuracy	Precision	Recall	F1
xlm-roberta-base	279M	0.9	0.88	0.88	0.88
bert-small-amharic	25.7M	0.89	0.86	0.87	0.86
bert-mini-amharic	9.67M	0.87	0.83	0.83	0.83

Original CSV and Paper

The original csv file can be found in this git repository https://github.com/IsraelAbebe/An-Amharic-News-Text-classification-Dataset

While there is a version of this dataset that's already available on huggingface hub (israel/Amharic-News-Text-classification-Dataset), that version had been preprocessed to remove punctuation from the articles, while this version contains the entire text along with punctuations. As a result, this version is more preferable for finetuning transformer models.

In NLP, text classification is one of the primary problems we try to solve and its uses in language analyses are indisputable. The lack of labeled training data made it harder to do these tasks in low resource languages like Amharic. The task of collecting, labeling, annotating, and making valuable this kind of data will encourage junior researchers, schools, and machine learning practitioners to implement existing classification models in their language. In this short paper, we aim to introduce the Amharic text classification dataset that consists of more than 50k news articles that were categorized into 6 classes. This dataset is made available with easy baseline performances to encourage studies and better performance experiments.

@misc{https://doi.org/10.48550/arxiv.2103.05639,
  doi = {10.48550/ARXIV.2103.05639},
  url = {https://arxiv.org/abs/2103.05639},
  author = {Azime, Israel Abebe and Mohammed, Nebil},
  keywords = {Computation and Language (cs.CL), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {An Amharic News Text classification Dataset},
  publisher = {arXiv},
  year = {2021},
  copyright = {arXiv.org perpetual, non-exclusive license}
}