A Knowledge-Aware Hybrid Summarization Framework with Cross-Domain Semantic Fusion, Entity-Aware Salience Scoring, and Hallucination-Suppressed Abstractive Generation for Biomedical Literature and Cli

Biomedical text summarization, hybrid summarization, cross-domain semantic fusion, entity-aware salience, hallucination suppression, multi-reward reinforcement learning, contrastive learning, transformer architectures, BART, BioBERT.

Siddu Tushara M S1*,
Research Scholar, JAIN (DEEMED-TO-BE UNIVERSITY)
Bengaluru, India.
tusharams23@gmail.com

Dr Renukadevi S2
Research Guide, JAIN (DEEMED-TO-BE UNIVERSITY)
Bengaluru, India.
renuka.devi@jainuniversity.ac.in

Received:15-06-2026
Revised:20-07-2026
Accepted: 24-07-2026

European Journal of Prosthodontics and Restorative Dentistry (2026) 34(5s), 314–331

A Knowledge-Aware Hybrid Summarization Framework with Cross-Domain Semantic Fusion, Entity-Aware Salience Scoring, and Hallucination-Suppressed Abstractive Generation for Biomedical Literature and Clinical Text

Due to the exponentially growing amount of biomedical literature with PubMed reaching more than 36 million indexed publications by 2025, the manual extraction of knowledge is operationally unachievable, thus necessitating the development of automated systems that produce clinically meaningful summaries. Three systematic shortcomings can be identified among state-of-theart neural summarization methods: knowledge-agnostic sentence selection failing to select the sentences containing PICO-structured evidence, abstractive decoder that is not restricted to generate entity-mentioning sentences, and the hybrid architectures that are not tightly integrated and, hence, hinder the exchange of knowledge between the extractive and abstractive modules. In this paper, we introduce BioSummHybrid+, an approach to Knowledge-Aware Hybrid Summarization that tackles all the above issues by integrating the following five innovations: Cross-Domain Semantic Fusion Gate (CSFG), Entity-Aware Salience Scorer (EASS), Hierarchical Discourse-Aware Positional Encoding (HDAPE), Multi-Reward Reinforcement Learning Decoding (MRLD), and Contrastive Hallucination Suppression (CHS). Our framework, BioSummHybrid+, was compared on five biomedical benchmarks (PubMed, SciTLDR, MS2, MultiXSci, ArXiv) against six baselines on two metrics (BERTScore-F1 and Faithfulness and Entity Preservation Rate) that reflect summarization quality in the context of high-stakes healthcare in a more adequate way compared to ROUGE and BLEU. It turns out that the higher ROUGE score is achieved by baselines through simple extractive lexical copying. Thus, we show that structured domain knowledge plays a vital role in summarization architecture.

•••••••••••••••••••••••••••••••• ejprd.org- Published by Riset Publishing Services LLC.

EJPRD

Copyright © 2026 by Riset Publishing Services LLC