arXiv cs.CL· Anas Belfathi, Nicolas Hernandez, Laura Monceaux, Warren Bonnard, Richard Dufour·· 13 小时前AI 评分8
RISE：面向修辞角色标注难例的推理时语义重排序框架
Semantic Reranking at Inference Time for Hard Examples in Rhetorical Role Labeling
AI 导读
RISE 是一个推理时语义重排序框架，自动识别低置信度预测，并用对比学习得到的标签表示对模型输出重排序，无需重训练或修改底层模型。在八个领域特定 RRL 数据集、七种语言模型（含编码器与因果架构）上，难例平均提升 +9.15 macro-F1。该论文已被作者撤回，原因是发现归属问题，请勿引用此版本。
正文 · 原文
This paper has been withdrawn by Anas Belfathi
No PDF available, click to view other formats
Abstract:Rhetorical Role Labeling (RRL) assigns a functional role to each sentence in a document and is widely used in legal, medical, and scientific domains. While language models (LMs) achieve strong average performance, they remain unreliable on hard examples, where prediction confidence is low. Existing approaches typically handle uncertainty implicitly and treat labels as discrete identifiers, overlooking the semantic information encoded in label names. We introduce RISE, an inference-time semantic reranking framework that leverages label semantics to refine predictions on hard instances. RISE automatically identifies low-confidence predictions and reranks model outputs using contrastively learned label representations, without retraining or modifying the underlying model. Experiments on eight domain-specific RRL datasets with seven LMs, including encoder-based and causal architectures, show an average gain of +9.15 macro-F1 points on hard examples. For explainability, we further propose manual hardness annotations to study difficulty from both model and human perspectives, revealing a moderate agreement with Cohen's kappa = 0.40.
| Comments: | This submission is being withdrawn by the authors following the identification of an attribution issue. The paper was withdrawn from the conference publication process before official publication. Please do not cite this version |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2605.18007 [cs.CL] |
| (or arXiv:2605.18007v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2605.18007 arXiv-issued DOI via DataCite |
Submission history
From: Anas Belfathi [view email]
[v1]
Mon, 18 May 2026 08:03:02 UTC (6,623 KB)
[v2]
Tue, 27 Oct 2026 15:10:56 UTC (1 KB) (withdrawn)
来源：arXiv cs.CL · arxiv.org