A Hypervolume-Based Comparison of NSGA-II and NSGA-III for Three-Objective Extractive Text Summarization

Long documents pose challenges both for human readers, who may struggle to quickly extract the main ideas, and for retrieval-augmented generation (RAG) systems, which face difficulties in selecting and providing the most relevant context when processing entire texts. Extractive text summarization represents a solution for this, but its usefulness depends on selecting the best possible summary considering, among other objectives, the essential information with minimal length. This work addresses focuses on identifying high-quality summaries by treating sentence selection as multi-objective optimization problem that must balance competing objectives: ensuring semantic coverage, prioritizing relevant portions of the text, and maintaining the summary sufficiently short. It compares two widely used optimization evolutionary algorithms, NSGA-II and NSGA-III, on 462 CNN DailyMail news articles. The results indicate a superiority of NSGA-II when analyzing quality alone, but NSGA-III presents an advantage when considering performance and solution selectivity, which enriches the discussion on the benefits trade-offs.