indonesia-news-classification-bert

Model Card for Indonesian News Classification Model

Model Description

This model is fine-tuned for the specific task of classifying Indonesian news articles (data were extracted from iqballx/indonesian_news_datasets) into predefined categories. It was trained using a dataset that was created by translating Indonesian news articles into English using a Neural Machine Translation (NMT) system and then labeling them with niksmer/ManiBERT, a model trained to classify political texts. The resulting dataset contains parallel corpora of Indonesian and English news texts alongside their corresponding categories.

Training Data

The training data consists of articles from the iqballx/indonesian_news_datasets which were translated to English and then labeled using the niksmer/ManiBERT model. The dataset includes various categories, capturing a wide array of topics.

Evaluation

The model was evaluated on a held-out test set, and its performance was measured in terms of accuracy. During the training process, the model's accuracy improved across multiple epochs, with the following accuracy scores achieved: 61.71% after the first epoch, 64.62% after the second epoch, 65.64% after the third epoch, and 65.27% after the fourth epoch. These results demonstrate the model's ability to consistently make correct classifications across different categories, indicating its robust performance.

Limitations and Bias

As with any machine learning model, it is important to recognize potential limitations and biases. The translation step could introduce errors or nuances that affect the labeling accuracy. Additionally, the ManiBERT model used for initial labeling was trained on political texts, which may limit its effectiveness on non-political news or introduce political bias.

How to Use the Model

To classify an Indonesian news article, you can use the script below:

from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_name = "YagiASAFAS/indonesia-news-classification-bert"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Write Indonesian Text
inputs = tokenizer("[Indonesian Text]", return_tensors="pt")

outputs = model(**inputs)
predictions = torch.nn.functional.softmax(outputs.logits, dim=1)

id2label = model.config.id2label

predicted_class_index = torch.argmax(predictions, dim=1).item()

predicted_class_index

predicted_category = id2label.get(predicted_class_index)

print("Predicted Category:", predicted_category)

Label Mapping