SwahBERT-base-cased

SwahBERT: Language model of Swahili

The model and all credits belong to the original authors. The model was uploaded to the HuggingFace Hub for convenience. For more details, please refer the original repository.

Is a pretrained monolingual language model for Swahili. <br> The model was trained for 800K steps using a corpus of 105MB that was collected from news sites, online discussion, and Wikipedia. <br> The evaluation was perfomed on several downstream tasks such as emotion classification, news classification, sentiment classification, and Named entity recognition.

import torch
from transformers import BertTokenizer

tokenizer = BertTokenizer.from_pretrained("swahbert-base-uncased")

# Tokenized input
text = "Mlima Kilimanjaro unapatikana Tanzania"
tokenized_text = tokenizer.tokenize(text)

SwahBERT => ['mlima', 'kilimanjaro', 'unapatikana', 'tanzania']
mBERT => ['ml', '##ima', 'ki', '##lima', '##nja', '##ro', 'una', '##patikana', 'tan', '##zania']

Pre-training data

The text was extracted from different sorces;<br>

• News sites: United Nations news, Voice of America (VoA), Deutsche Welle (DW) and taifaleo<br>
• Forums: JaiiForum<br>
• Wikipedia.

Pre-trained Models

Download the models here:<br>

• SwahBERT-Base, Uncased:12-layer, 768-hidden, 12-heads , 124M parameters
• SwahBERT-Base, Cased:12-layer, 768-hidden, 12-heads , 111M parameters

Emotion Dataset

We released the Swahili emotion dataset.<br> The data consists of ~13K emotion annotated comments from social media platforms and translated English dataset. <br> The data is multi-label with six Ekman’s emotions: happy, surprise, sadness, fear, anger, and disgust or neutral.

Evaluation

The model was tested on four downstream tasks including our new emotion dataset

F1-score of language models on downstream tasks

Citation

Please use the following citation if you use the model or dataset:

@inproceedings{martin-etal-2022-swahbert,
    title = "{S}wah{BERT}: Language Model of {S}wahili",
    author = "Martin, Gati  and Mswahili, Medard Edmund  and Jeong, Young-Seob  and Woo, Jiyoung",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.naacl-main.23",
    pages = "303--313"
    }