PEGASUS-Based Text Summarization

This chapter introduces the implementation of Chinese text summarization using the PEGASUS pre-trained model within PaddleNLP. It begins with an overview of extractive and generative summarization methods, highlighting the advantages of generative approaches enabled by encoder–decoder architectures. The chapter then details the use of the LCSTS dataset, data preprocessing steps, feature construction, and DataLoader preparation. The PEGASUS model is trained with AdamW optimization and evaluated using BLEU and ROUGE metrics, demonstrating steadily improving performance with training. Experimental results show strong summarization quality, with high ROUGE and BLEU scores on the test set. Finally, the chapter presents model inference examples, practical applications, and exercises for extending and improving summarization approaches.