[2406.04625] Key-Element-Informed sLLM Tuning for Document Summarization

View a PDF of the paper titled Key-Ingredient-Knowledgeable sLLM Tuning for Doc Summarization, by Sangwon Ryu and 4 different authors

View PDF
HTML (experimental)

Summary:Outstanding advances in giant language fashions (LLMs) have enabled high-quality textual content summarization. Nevertheless, this functionality is presently accessible solely by means of LLMs of considerable measurement or proprietary LLMs with utilization charges. In response, smaller-scale LLMs (sLLMs) of simple accessibility and low prices have been extensively studied, but they usually endure from lacking key info and entities, i.e., low relevance, particularly, when enter paperwork are lengthy. We therefore suggest a key-element-informed instruction tuning for summarization, so-called KEITSum, which identifies key components in paperwork and instructs sLLM to generate summaries capturing these key components. Experimental outcomes on dialogue and information datasets reveal that sLLM with KEITSum certainly offers high-quality summarization with increased relevance and fewer hallucinations, aggressive to proprietary LLM.