---
license: apache-2.0
language:
- sw
---
# 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](https://github.com/gatimartin/SwahBERT)**.

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.

```ruby
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`](https://drive.google.com/drive/folders/1HZTCqxt93F5NcvgAWcbrXZammBPizdxF?usp=sharing)**:12-layer, 768-hidden, 12-heads , 124M parameters
- **[`SwahBERT-Base, Cased`](https://drive.google.com/drive/folders/1cCcPopqTyzY6AnH9quKcT9kG5zH7tgEZ?usp=sharing)**:12-layer, 768-hidden, 12-heads , 111M parameters

Steps | vocab size | MLM acc | NSP acc | loss |
--- | --- | --- | --- | --- |
**800K** | **50K (uncased)** | **76.54** | **99.67** | **1.0667** |
**800K** | **32K (cased)** | **76.94** | **99.33** | **1.0562** | 


## Emotion Dataset
We released the **[`Swahili emotion dataset`](https://github.com/gatimartin/SwahBERT/tree/main/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 
Tasks | SwahBERT | SwahBERT_cased | mBERT |                        
--- | --- | --- | --- |
Emotion | 64.46 | 64.77 | 60.52 | 
News | 90.90 | 89.90 | 89.73 |
Sentiment | 70.94 | 71.12 | 67.20 |
NER | 88.50 | 88.60 | 89.36 |


## 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"
    }
```