An Explainable Hybrid Deep Learning Framework for Outcome-Oriented Long-Document Summarization Using Adaptive Feature Learning and Decision Optimization

Abstract Automatic long-document summarization has seen great success in terms of transformer-based architecture and LLMs; however, the current techniques focus mainly on semantic relevance rather than alignment with learning objectives in educational settings. In this work, we introduce an Explainable Hybrid Deep Learning Framework that consists of hybrid semantic retrieval, parameter-efficient transformer fine-tuning, and alignment-oriented optimization to create outcome-oriented instructional summaries. Specifically, we perform dense retrieval with cross-encoder reranking to extract relevant evidence and then use an LED-based summarization model with LoRA for efficient long-context learning. The proposed method is tested on three different educational datasets - QMSum, GovReport, and NILD - and lexical, semantic, and outcome-oriented evaluation metrics are used. As the experimental results show, our technique outperforms state-of-the-art approaches significantly and obtains the following ROUGE-1 scores (40.01, 58.27, 53.57), BERTScores (88.12%, 81.27%, 87.27%), and Outcome Alignment Scores (OAS) (90.02%, 89.65%, 93.04%) on corresponding datasets.