Enhancing Text Summarization Capability of Lightweight Models through Dynamic Direct Preference Optimization(DPO) Mechanism

ACL ARR 2025 May Submission3785 Authors

Abstract: The abstractive summarization is a natural language processing(NLP) task that involves generating concise summaries of longer documents while preserving key information. Currently, state-of-art summarization methods are dominated by large language models (LLMs), their strong understandings, and generalizations have reshaped summarization research. Unlike those works, we focus on developing a light yet efficient abstractive summarizer targeting for edge-device applications. The primary challenge lies in the limited context understanding and paraphrasing abilities of lightweight models, constrained by their smaller capacity and vocabulary size. To address this, we introduce a novel framework integrating an online feedback mechanism. This system incorporates improvement suggestions to dynamically adjust the model’s outputs, enhancing its learning capabilities. Our approach achieves state-of-the-art (SOTA) results on CNN/DailyMail and XSum, outperforming backbones by 19.3% and 12.9%, respectively.

Paper Type: Short

Research Area: Summarization

Research Area Keywords: Text Summarization; Lightweight models;

Contribution Types: NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 3785