Reproduce “TriSum: Learning Summarization Ability from Large Language Models with Structured Rationale” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pengcheng Jiang, Cao Xiao, Zifeng Wang, Parminder Bhatia, Jimeng Sun, Jiawei Han
A three-step framework, TriSum, effectively distills the structured summarization and reasoning capabilities of large language models into smaller, local models. This method yields improved ROUGE scores and factual consistency across diverse datasets, providing more interpretable summaries than existing baselines.
Key insights
- Explicitly integrating LLM-generated structured rationales (aspects and knowledge graph-like triples) serves as effective additional supervision, enhancing the performance and interpretability of smaller summarization models.
- A dual-scoring mechanism for selecting 'golden' rationales, considering both summary quality and topical coherence with the document, is crucial for distilling high-quality reasoning from LLMs.
- A phased curriculum learning approach, starting with basic task understanding before progressing to complex joint generation, is essential for a compact model to successfully learn and leverage LLM-derived structured reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.