---
language: su
tags:
  - sundanese-gpt2-base-emotion-classifier
license: mit
widget:
  - text: "Wah, éta gélo, keren pisan!"
---

## Sundanese GPT-2 Base Emotion Classifier

Sundanese GPT-2 Base Emotion Classifier is an emotion-text-classification model based on the [OpenAI GPT-2](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) model. The model was originally the pre-trained [Sundanese GPT-2 Base](https://hf.co/w11wo/sundanese-gpt2-base) model, which is then fine-tuned on the [Sundanese Twitter dataset](https://github.com/virgantara/sundanese-twitter-dataset), consisting of Sundanese tweets.

10% of the dataset is kept for evaluation purposes. After training, the model achieved an evaluation accuracy of 94.84% and F1-macro of 94.75%.

Hugging Face's `Trainer` class from the [Transformers](https://huggingface.co/transformers) library was used to train the model. PyTorch was used as the backend framework during training, but the model remains compatible with other frameworks nonetheless.

## Model

| Model                                    | #params | Arch.      | Training/Validation data (text) |
| ---------------------------------------- | ------- | ---------- | ------------------------------- |
| `sundanese-gpt2-base-emotion-classifier` | 124M    | GPT-2 Base | Sundanese Twitter dataset       |

## Evaluation Results

The model was trained for 10 epochs and the best model was loaded at the end.

| Epoch | Training Loss | Validation Loss | Accuracy | F1       | Precision | Recall   |
| ----- | ------------- | --------------- | -------- | -------- | --------- | -------- |
| 1     | 0.819200      | 0.331463        | 0.880952 | 0.878694 | 0.883126  | 0.879304 |
| 2     | 0.140300      | 0.309764        | 0.900794 | 0.899025 | 0.906819  | 0.898632 |
| 3     | 0.018600      | 0.324491        | 0.948413 | 0.947525 | 0.948037  | 0.948153 |
| 4     | 0.004500      | 0.335100        | 0.932540 | 0.931648 | 0.934629  | 0.931617 |
| 5     | 0.000200      | 0.392145        | 0.932540 | 0.932281 | 0.935075  | 0.932527 |
| 6     | 0.000000      | 0.371689        | 0.932540 | 0.931760 | 0.934925  | 0.931840 |
| 7     | 0.000000      | 0.368086        | 0.944444 | 0.943652 | 0.945875  | 0.943843 |
| 8     | 0.000000      | 0.367550        | 0.944444 | 0.943652 | 0.945875  | 0.943843 |
| 9     | 0.000000      | 0.368033        | 0.944444 | 0.943652 | 0.945875  | 0.943843 |
| 10    | 0.000000      | 0.368391        | 0.944444 | 0.943652 | 0.945875  | 0.943843 |

## How to Use

### As Text Classifier

```python
from transformers import pipeline

pretrained_name = "sundanese-gpt2-base-emotion-classifier"

nlp = pipeline(
    "sentiment-analysis",
    model=pretrained_name,
    tokenizer=pretrained_name
)

nlp("Wah, éta gélo, keren pisan!")
```

## Disclaimer

Do consider the biases which come from both the pre-trained RoBERTa model and the Sundanese Twitter dataset that may be carried over into the results of this model.

## Author

Sundanese GPT-2 Base Emotion Classifier was trained and evaluated by [Wilson Wongso](https://w11wo.github.io/). All computation and development are done on Google Colaboratory using their free GPU access.

## Citation Information

```bib
@article{rs-907893,
    author   = {Wongso, Wilson
                and Lucky, Henry
                and Suhartono, Derwin},
    journal  = {Journal of Big Data},
    year     = {2022},
    month    = {Feb},
    day      = {26},
    abstract = {The Sundanese language has over 32 million speakers worldwide, but the language has reaped little to no benefits from the recent advances in natural language understanding. Like other low-resource languages, the only alternative is to fine-tune existing multilingual models. In this paper, we pre-trained three monolingual Transformer-based language models on Sundanese data. When evaluated on a downstream text classification task, we found that most of our monolingual models outperformed larger multilingual models despite the smaller overall pre-training data. In the subsequent analyses, our models benefited strongly from the Sundanese pre-training corpus size and do not exhibit socially biased behavior. We released our models for other researchers and practitioners to use.},
    issn     = {2693-5015},
    doi      = {10.21203/rs.3.rs-907893/v1},
    url      = {https://doi.org/10.21203/rs.3.rs-907893/v1}
}
```