turkish-news-classifier

turkish-news-classifier

A fine-tuned version of dbmdz/bert-base-turkish-cased for multi-class news category classification of Turkish text.

Model description

The model takes a Turkish news article (or headline) and predicts one of twelve category labels. It is a standard BertForSequenceClassification head fine-tuned on top of the cased Turkish BERT (BERTurk) encoder.

• Base model: dbmdz/bert-base-turkish-cased
• Language: Turkish (tr)
• Task: Single-label, multi-class text classification (news category)
• Number of classes: 12

Labels

The label ids follow model.config.id2label:

Intended uses & limitations

Intended use: classifying the primary topic of Turkish news articles or headlines (e.g. for tagging, routing, or filtering a news feed).

The model expects raw text — no special preprocessing is required. Just pass the headline and/or article body directly; the tokenizer (shipped with the model) handles everything.

Limitations / out of scope:

• Very long documents get truncated at the model's max sequence length; for long articles, consider passing just the headline + lead paragraph.
• The training data is Turkish news from roughly 2010–2013, so the model may not reflect current events, entities, or writing style.
• Category boundaries in the source corpus can overlap (e.g. genel/guncel/turkiye cover similar "general domestic news" ground) — this is a property of the labeling scheme, not something the model can resolve on its own.
• Not intended for languages other than Turkish or for categories outside the trained set.

How to use

from transformers import pipeline

pipe = pipeline("text-classification", model="furkankarakuz/turkish-news-classifier")
print(pipe("Galatasaray, derbi maçında rakibini 3-1 mağlup etti."))
# [{'label': 'spor', 'score': 0.97}]

Or load the model and tokenizer directly:

from transformers import AutoModelForSequenceClassification, AutoTokenizer

model = AutoModelForSequenceClassification.from_pretrained("furkankarakuz/turkish-news-classifier")
tokenizer = AutoTokenizer.from_pretrained("furkankarakuz/turkish-news-classifier")

Training and evaluation data

<!-- TODO: fill in the exact sample count and split sizes from your actual prepare_data.py / notebook run. -->

Trained on the 42000 Turkish News Texts in 13 Classes corpus (YTÜ KEMİK NLP Group), approximately 40,000 raw Turkish news articles. The original corpus's planet category was dropped before training — its content overlapped with dunya (both cover world/international news) — leaving the 12 categories listed above. The data was split with stratification so the class balance is preserved across splits:

• Train: 70%
• Validation: 15%
• Test: 15%

Labels were integer-encoded with LabelEncoder, fit on the training split only. No lossy text cleaning was applied beyond whitespace normalization; the model is trained on raw article text. Classes are imbalanced (spor is the largest category, kultur-sanat/saglik among the smallest), so training used a class-weighted loss (see below).

Training procedure

Fine-tuned with the 🤗 Transformers Trainer, using a class-weighted cross-entropy loss to handle class imbalance. Training ran for 8 epochs (early stopping, patience 2) before stopping; the metrics below correspond to that final checkpoint.

Training hyperparameters

Validation results per epoch

Training stopped after epoch 8 (early stopping, patience 2). Note validation macro-F1 peaked slightly earlier, at epoch 6 (0.7056) — epoch 8 is shown as final since that's the checkpoint that was evaluated and pushed.

Limitations and bias

The model reflects the distribution and any biases of its training data (Turkish news from ~2010–2013) and may underperform on current events, entities, or writing styles not well represented in it. Macro-F1 (~0.70) is noticeably lower than typical short-text classification tasks (e.g. emotion classification on this same base model reaches ~0.97), reflecting genuine ambiguity and overlap between some of the 12 news categories (e.g. genel vs. guncel vs. turkiye) rather than a modeling shortfall alone. Predictions should be treated as a "most likely category" signal rather than a ground-truth label, especially near these overlapping category boundaries.

Framework versions

• Transformers 5.13.1
• PyTorch 2.11.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2