arXiv cs.AI· Aarush Sinha, Rahul Seetharaman, Aman Bansal·· 2 天前AI 评分18
什么在追踪难负样本迁移？词汇与语义信号研究
What Tracks Hard-Negative Transfer? A Study of Lexical and Semantic Signals
AI 导读
一项研究比较了词汇非覆盖度、冻结编码器相似度统计、难度代理等信号，在 BM25、稠密、LLM 生成及混合来源的 MS MARCO 训练数据上对稠密检索器迁移的追踪能力。在三个检索器骨干和十一个 BEIR 数据集上，IDF 加权的平均词汇残差性复现了全部六组骨干与设置块的迁移排序，而语义统计量均无法恢复完整排序，与词汇信号结合也未提升排序一致性。
正文 · 原文
Abstract:Which properties of hard-negative supervision track dense-retriever transfer before training? We compare lexical noncoverage, frozenencoder similarity statistics, hardness proxies, and combinations of these signals across BM25, dense, LLM-generated, and hybrid sources for MS MARCO training. Evaluation uses three retriever backbones and eleven BEIR datasets. Mean lexical residuality, an IDF-weighted measure of query-token noncoverage, reproduces the aggregate transfer ordering in all six backbone and setting blocks. The individual semantic statistics recover none of the complete orderings, and combining them with the lexical signal does not improve the observed rank agreement. After approximately matching BM25, dense, and LLM diagnostic samples for lexical residuality, none of the tested semantic statistics recovers the observed LLM > Dense > BM25 transfer ordering across three backbones, three matching resolutions, and 100 repetitions. Constructed controls further show that unrelated and false-negative-shaped passages can exhibit favorable lexical and semantic statistics. The findings identify a simple lexical association in this setting while distinguishing source level transfer agreement from the reliability of individual negatives.
| Subjects: | Information Retrieval (cs.IR); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2603.20990 [cs.IR] |
| (or arXiv:2603.20990v5 [cs.IR] for this version) | |
| https://doi.org/10.48550/arXiv.2603.20990 arXiv-issued DOI via DataCite |
Submission history
From: Aarush Sinha [view email]
[v1]
Sun, 22 Mar 2026 00:21:05 UTC (71 KB)
[v2]
Wed, 3 Jun 2026 21:47:20 UTC (59 KB)
[v3]
Fri, 5 Jun 2026 15:52:58 UTC (60 KB)
[v4]
Tue, 11 Aug 2026 12:10:27 UTC (573 KB)
[v5]
Wed, 14 Oct 2026 10:06:02 UTC (53 KB)
来源：arXiv cs.AI · arxiv.org