ECSum: Evidence-Constrained and Preference-Aligned Text Summarization for Veterinary and Animal Science Literature

Abstract

Transformer-based summarizers like BART, T5, PEGASUS, and long-sequence variants provide strong baseline for abstractive summarization tasks, but they still have problems with long-context understanding and factual consistency is not good enough. Standard metric such as ROUGE often not align well with what expert actually think about summary quality [1–4], [5–8]. In veterinary medicine field and animal science research area, making accurate summaries of clinical case reports, livestock production study, and wildlife health surveillance document is very critical thing—if model generate hallucinated finding, it can cause misdiagnosis or improper decision for herd management which is dangerous. We propose ECSum, a unified framework that combine: (i) planner–verifier mechanism for long input using Transient-Global attention, (ii) evidence-coverage loss and entailment regularization during training process for reducing hallucination problem, and (iii) evidence-aware preference optimization using Direct Preference Optimization (DPO) to make output align with what veterinarian and animal scientist prefer. We conduct evaluations across veterinary case reports, livestock research documents, poultry science paper, and wildlife health monitoring data with hybrid protocol that combine ROUGE/BERTScore, QA-based factuality measure (QAFactEval), and LLM-assisted meta-evaluation using G-Eval. Results show ECSum consistently improve factual consistency and expert preference while achieving competitive or better ROUGE score compared to BART-large, PEGASUS, and LongT5; statistical test indicate significance on most datasets we test. Main contributions of this work include: (1) training-time evidence-constrained objective for long-document summarization in animal science domain; (2) evidence-aware reward model for preference alignment that veterinary professional validate; and (3) reproducible hybrid evaluation pipeline that reflect domain expert judgment better than before.