ECR: An Expertise-Enriched Conclude-Then-Refine Summarization Framework for Professional Articles

Summary generation using large language models (LLMs) is characterized by its flexibility, high quality, and efficiency. However, professional articles usually contain domain-specific background knowledge and many professional terminologies. It’s hard for the generated summary to maintain professionalism and good writing style using simple prompts coupled with LLMs, which is the common summarization method. While developing task-specific LLMs can improve the summary quality, it demands a high training cost. To enhance the summary quality cost-efficiently, we present ECR, a two-stage expertise-enriched conclude-then-refine summarization framework for professional articles. Firstly, the Key Information Conclusion Stage (KICS) distills the article content through elaborated prompt engineering. Subsequently, the Refinement and Term Enrichment Stage (RTES) enhances coherence, conciseness, and professionalism. Experimental results indicate that our approach offers superior summaries to the summaries generated by the common method across diverse domains. ECR shows an over 80% win rate in structural consistency evaluated by the LLM, alongside a 2x increase in terminology integration. The framework also provides sufficient flexibility to replace components within it. Furthermore, a test dataset comprising 100 articles and an evaluation method for assessing professionalism are proposed.