arXiv cs.CL· Maitreya Prafulla Chitale, Ketaki Mangesh Shetye, Yash More, Harshit Gupta, Manav Chaudhary, Manish Shrivastava, Vasudeva Varma·· 2 天前AI 评分20
单一参考还是多个参考？重新思考中等熵文本生成的多参考训练
One Reference or Many? Rethinking Multi-Reference Training for Intermediate-Entropy Text Generation
AI 导读
针对同行评审生成和图像描述等存在多个有效输出的中等熵文本生成任务，研究者提出目标端整合方法，将同一输入的所有参考合并为保留全部独立信息单元的单一目标，并构建了包含36,537篇论文和31,014张图像的MERC-67K数据集。
正文 · 原文
Abstract:Natural language generation (NLG) ranges from highly constrained tasks, such as machine translation, to open-ended tasks like dialogue generation. Between them lie peer-review generation and image captioning, where one input admits several valid, overlapping outputs. We measure this intermediary region with Von Neumann entropy ($\hat{H}\approx0.6$ vs. 0.15 translation, 0.92 dialogue). Single-reference training discards valid perspectives, while multi-reference training introduces high-variance, unaligned targets. To tackle this, we propose target-side consolidation, that is, merging all references for an input into a single target which preserves every distinct information unit. To this end, we introduce MERC-67K, pairing 36,537 papers and 31,014 images with their original and consolidated references. In human evaluation, models trained using consolidated reviews obtain an average rank of 1.70 (out of 4) for coverage against 2.51-2.98 for baselines, and win 63.5-74.0% of pairwise comparisons. Models trained using consolidated captions obtain an average rank of 1.87 for overall caption quality (accuracy and naturalness) against 2.64-2.82 for baselines and obtain the highest reference-free LongCLIPScore (0.839 vs best baseline's 0.816). For reviews, multi-reference training often under-performs zero-shot inference, and consolidation's gains over multi-reference training hold across Llama3.1, Mistralv0.3, and Qwen2.5. Consolidation uses $\sim$4$\times$ (reviews) and 5$\times$ (captions) fewer individual references during optimization. Dataset and code will be released upon acceptance.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2609.22805 [cs.CL] |
| (or arXiv:2609.22805v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2609.22805 arXiv-issued DOI via DataCite |
Submission history
From: Maitreya Chitale [view email]
[v1]
Sat, 19 Sep 2026 06:13:28 UTC (1,138 KB)
[v2]
Thu, 15 Oct 2026 08:41:02 UTC (631 KB)
来源：arXiv cs.CL · arxiv.org