File size: 3,607 Bytes
1a80774 9892ae1 2d9a5d7 9892ae1 4c5d853 9892ae1 2d9a5d7 9892ae1 cae6c4b 9892ae1 398b249 9892ae1 cf3deee 9892ae1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 |
---
annotations_creators:
- crowdsourced
language:
- amh
- hau
- ibo
- kin
- lug
- luo
- pcm
- swa
- wol
- yor
- bam
- bbj
- ewe
- fon
- mos
- nya
- sna
- tsn
- twi
- xho
- zul
language_creators:
- crowdsourced
license:
- cc-by-4.0
multilinguality:
- monolingual
pretty_name: afrolm-dataset
size_categories:
- 1M<n<10M
source_datasets:
- original
tags:
- afrolm
- active learning
- language modeling
- research papers
- natural language processing
- self-active learning
task_categories:
- fill-mask
task_ids:
- masked-language-modeling
---
# AfroLM: A Self-Active Learning-based Multilingual Pretrained Language Model for 23 African Languages
- [GitHub Repository of the Paper](https://github.com/bonaventuredossou/MLM_AL)
This repository contains the model for our paper `AfroLM: A Self-Active Learning-based Multilingual Pretrained Language Model for 23 African Languages` which will appear at the Third Simple and Efficient Natural Language Processing, at EMNLP 2022.
## Our self-active learning framework
## Languages Covered
AfroLM has been pretrained from scratch on 23 African Languages: Amharic, Afan Oromo, Bambara, Ghomalá, Éwé, Fon, Hausa, Ìgbò, Kinyarwanda, Lingala, Luganda, Luo, Mooré, Chewa, Naija, Shona, Swahili, Setswana, Twi, Wolof, Xhosa, Yorùbá, and Zulu.
## Evaluation Results
AfroLM was evaluated on MasakhaNER1.0 (10 African Languages) and MasakhaNER2.0 (21 African Languages) datasets; on text classification and sentiment analysis. AfroLM outperformed AfriBERTa, mBERT, and XLMR-base, and was very competitive with AfroXLMR. AfroLM is also very data efficient because it was pretrained on a dataset 14x+ smaller than its competitors' datasets. Below is the average performance of various models, across various datasets. Please consult our paper for more language-level performance.
Model | MasakhaNER | MasakhaNER2.0* | Text Classification (Yoruba/Hausa) | Sentiment Analysis (YOSM) | OOD Sentiment Analysis (Twitter -> YOSM) |
|:---: |:---: |:---: | :---: |:---: | :---: |
`AfroLM-Large` | **80.13** | **83.26** | **82.90/91.00** | **85.40** | **68.70** |
`AfriBERTa` | 79.10 | 81.31 | 83.22/90.86 | 82.70 | 65.90 |
`mBERT` | 71.55 | 80.68 | --- | --- | --- |
`XLMR-base` | 79.16 | 83.09 | --- | --- | --- |
`AfroXLMR-base` | `81.90` | `84.55` | --- | --- | --- |
- (*) The evaluation was made on the 11 additional languages of the dataset.
- Bold numbers represent the performance of the model with the **smallest pretrained data**.
## Pretrained Models and Dataset
**Models:**: [AfroLM-Large](https://huggingface.co/bonadossou/afrolm_active_learning) and **Dataset**: [AfroLM Dataset](https://huggingface.co/datasets/bonadossou/afrolm_active_learning_dataset)
## HuggingFace usage of AfroLM-large
```python
from transformers import AutoTokenizer, AutoModelForTokenClassification
model = AutoModelForTokenClassification.from_pretrained("bonadossou/afrolm_active_learning")
tokenizer = AutoTokenizer.from_pretrained("bonadossou/afrolm_active_learning")
tokenizer.model_max_length = 256
```
## Reproducing our result: Training and Evaluation
- To train the network, run `python active_learning.py`. You can also wrap it around a `bash` script.
- For the evaluation:
- NER Classification: `bash ner_experiments.sh`
- Text Classification & Sentiment Analysis: `bash text_classification_all.sh`
## Citation
We will share the proceeding citation as soon as possible. Stay tuned, and if you have liked our work, give it a star.
## Reach out
Do you have a question? Please create an issue and we will reach out as soon as possible |