Multilingual Abstractive Summarization With Mt5 Encoder Decoder Architecture

via “abstractive-summarization-with-bart-encoder-decoder”

summarization model by undefined. 19,66,142 downloads.

Unique: Uses BART's denoising autoencoder architecture (trained with corrupted input reconstruction) combined with CNN/DailyMail fine-tuning, enabling abstractive summarization that generates novel phrasings rather than extractive copying. The encoder-decoder design with cross-attention allows the model to dynamically attend to relevant source passages while generating each summary token, unlike simpler seq2seq models.

vs others: Outperforms extractive summarization baselines and earlier seq2seq models on ROUGE metrics for news summarization; more abstractive than PEGASUS but with faster inference than T5-large due to smaller parameter count (406M vs 770M), making it the practical choice for resource-constrained production deployments.