HLAS: An Enhanced Abstractive Summarization Method Based on Hybrid Loss Optimization

The conventional training of summarization models using maximum likelihood loss faces a challenge: the discrepancy between training and evaluation objectives. Training minimizes token-level loss, while evaluation focuses on sentence or document quality, leading to less effective results. To address this, we propose the Hybrid Loss Abstractive Summarization (HLAS) framework, which combines maximum likelihood loss with global sequence prediction metrics, aligning training with evaluation. Experiments on the BillSum and GovReport datasets show that HLAS significantly outperforms baseline models, achieving higher ROUGE scores and better summarization quality.