mizgithub/Amharic_news_headline_generation

# Fine-Tuned Pretrained Transformer for Amharic News Headline Generation This repository contains the code and resources for the paper **"Fine-Tuned Pretrained Transformer for Amharic News Headline Generation"**. The project aims to generate news headlines for Amharic articles using a **fine-tuned T5v1.1 model (t5-small)**. The model was trained on **over 70k Amharic news articles** and evaluated using **ROUGE-L, BLEU, and METEOR metrics**, achieving competitive performance compared to non-fine-tuned models and previous studies. ## 🔗 Paper Publication **Fine-Tuned Pretrained Transformer for Amharic News Headline Generation**. Link to the Paper ## 📄 Abstract Amharic is an under-resourced language, which makes automatic news headline generation a challenging task due to the limited availability of high-quality datasets. This project fine-tunes the **T5v1.1 model (t5-small)** to generate Amharic news headlines using **TF-IDF optimization** and **Byte Pair Encoding (BPE) tokenization**. The system achieved: - **ROUGE-L**: 0.72 - **BLEU**: 0.52 - **METEOR**: 0.81 These results significantly outperform the baseline **non-fine-tuned T5 model**, which achieved ROUGE-L: 0.1, BLEU: 0.03, and METEOR: 0.14. The contributions of this study provide insights for further improvements, such as increasing dataset size, exploring other transformer models, and developing **adaptive post-processing techniques**. --- ## 🚀 How to Use the Code ### Step 1: Clone the Repository ```bash git clone github.com cd Amharic-news-headline-generation pip install -r requirements.txt ``` --- ### Step 2: Creating tokenizer, Optional. You can also use already created tokenizer #### Create tokenizer ``` from tokenization import custom_tokenizer tokenizer = custom_tokenizer(dataset) ``` ### Step3: Train or use the model ``` from amharic_nhg import ANHG #Initialize the model model = ANHG() # Train the model dataset = "path/to/dataset.csv" …