News Sum Tr

🚀 Turkish News Summarization Model Based on mT5

This model is designed to summarize Turkish news texts. It leverages the power of the mT5 architecture, specifically fine - tuned for the task of summarizing Turkish news. Built upon the google/mt5 - base architecture, it is trained on a dataset composed of Turkish economic and current news texts, aiming to provide users with the most crucial points of the news by generating summaries.

🚀 Quick Start

Basic Usage

from transformers import MT5Tokenizer, MT5ForConditionalGeneration

# Load the model and tokenizer
model_name = 'nebiberke/news - sum - tr'
tokenizer = MT5Tokenizer.from_pretrained(model_name)
model = MT5ForConditionalGeneration.from_pretrained(model_name)

# Generate a summary for a news text
news_text = "Buraya Türkçe haber metni gelecek."
input_ids = tokenizer.encode(news_text, return_tensors="pt", max_length=512, truncation=True)

# Generate the summary
summary_ids = model.generate(input_ids, max_length=150, min_length=30, length_penalty=2.0, num_beams=7, early_stopping=True)
summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)

print("Summary:", summary)

✨ Features

• Summarization: The primary function of this model is to transform long Turkish news texts into short and understandable summaries.
• Potential Applications: It can be used in news summarization platforms.

📦 Installation

No specific installation steps are provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

from transformers import MT5Tokenizer, MT5ForConditionalGeneration

# Load the model and tokenizer
model_name = 'nebiberke/news - sum - tr'
tokenizer = MT5Tokenizer.from_pretrained(model_name)
model = MT5ForConditionalGeneration.from_pretrained(model_name)

# Generate a summary for a news text
news_text = "Buraya Türkçe haber metni gelecek."
input_ids = tokenizer.encode(news_text, return_tensors="pt", max_length=512, truncation=True)

# Generate the summary
summary_ids = model.generate(input_ids, max_length=150, min_length=30, length_penalty=2.0, num_beams=7, early_stopping=True)
summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)

print("Summary:", summary)

Advanced Usage

There are no advanced usage examples in the original document, so this part is not added.

📚 Documentation

Model Information

This model is based on mT5 and is fine - tuned specifically for summarizing Turkish news texts. It is built on the google/mT5 - base architecture and trained on a combined dataset of batubayk/TR - News and denizzhansahin/100K - TR - News.

Model Purpose

The main purpose of this model is to convert long Turkish news texts into short and understandable summaries. Potential use cases include news summarization platforms.

Dataset

The model is trained on a combined dataset of batubayk/TR - News and denizzhansahin/100K - TR - News. The dataset contains Turkish economic and current news content, where each news text is paired with a summary.

• Training and Test Split: The dataset is split into 80% for training and 20% for testing.
• Maximum Input Length: 512 tokens
• Maximum Output Length (Summary): 150 tokens

Training Details

• Model Architecture: MT5 - base
• Training Objective: Seq2Seq summarization
• Batch Size: 4
• Number of Epochs: 6
• Weight Decay: 0.01
• Learning Rate: 5e - 5
• Evaluation Strategy: Evaluation is conducted at the end of each epoch.
• Logging: Logging is done at the end of each epoch.
• Device: CUDA - enabled GPU

Training and Evaluation Losses

• At the end of the first epoch, the training loss is 0.6440, and the validation loss is 0.5474.
• At the end of the last epoch, the training loss is 0.4879, and the validation loss is 0.4721.

Evaluation Results

The performance of the model is evaluated using ROUGE and BERTScore metrics.

ROUGE Scores:

• ROUGE - 1: Precision: 0.4231, Recall: 0.5000, F1 - score: 0.4583
• ROUGE - 2: Precision: 0.2400, Recall: 0.2857, F1 - score: 0.2609
• ROUGE - L: Precision: 0.4231, Recall: 0.5000, F1 - score: 0.4583

BERTScore:

• Precision: 0.9031
• Recall: 0.9196
• F1 - score: 0.9113

Limitations and Biases

• The model may learn biases present in the training data and reflect them in the summaries.
• A decline in summarization quality may be observed, especially for very long and complex news texts.

📄 License

No license information is provided in the original document, so this section is skipped.

📚 Additional Information

📖 Citation

If you use this model, please cite it as follows:

@article{2024turkishsummarization,
  title={mT5 Tabanlı Türkçe Haber Özetleme Modeli},
  author={Nebi Berke İçöz},
  year={2024},
  url={https://huggingface.co/nebiberke/news - sum - tr}
}