Abstractive Text Summarization with T5 and BART: Generating Concise News Summaries

Introduction: In today’s fast-paced world, reading full news articles is time-consuming. Abstractive text summarization helps condense long articles into meaningful summaries, capturing the essence without copying the original text.

Dataset: I used the CNN/DailyMail dataset containing news articles paired with human-written summaries. This dataset is ideal for training Encoder–Decoder models for abstractive summarization.

Model: I fine-tuned T5 / BART using Hugging Face Transformers. These models are encoder-decoder architectures designed to generate high-quality text summaries. Tokenization and padding were applied to ensure uniform input lengths.

Training: The model was trained with article-summary pairs, using a cross-entropy loss on the target summaries. The GPU-enabled training allowed faster iterations.

Evaluation: I evaluated the model using ROUGE metrics, which measure overlap between generated and reference summaries.

Metric Score

ROUGE-1 F1 = 0.369

ROUGE-2 F1 = 0.115

ROUGE-L F1 0.274

Qualitative Examples:

• Original: “The company reported strong earnings this quarter with profits increasing by 20%…”
• Reference: “Company profits grew 20% due to new products and market expansion.”
• Generated: “The company reported strong earnings this quarter with profits increasing by 20%. CEO expressed optimism about future growth prospects. The company’s profits were driven by successful product launches and expanding market share.”

Deployment:

I deployed the model using Streamlit, enabling real-time summarization. Users can input articles and instantly get concise summaries.