揭示交叉编码器如何模拟BM25实现语义相关性判断
Pathway to Relevance: How Cross-Encoders Implement a Semantic Variant of BM25
- 通过分析模型各层，发现早期层提取词频、逆文档频率等传统信号
- 后期层将这些信号组合成类似BM25的概率排序机制
- 为提升检索模型透明度与安全性提供可干预的结构基础
机制解释已显著促进对生成式语言模型内部结构的理解，使我们能够识别关键行为的实现机制。相比之下，信息检索（IR）领域的可解释性仍较粗略，关于IR模型如何判断文档与查询的相关性仍不明确。本文通过机制分析，研究了常用模型——交叉编码器——的相似性判断过程。结果表明，该模型在早期到中期层中提取了传统的相关性信号，如词频和逆文档频率；随后在后期层中将这些信号组合，形成类似经典概率排序函数BM25的机制。整体而言，本研究深化了对检索模型如何计算相关性的理解，分离出的关键组件为未来增强透明度、缓解安全风险及提升可扩展性的干预提供了基础。
原文摘要 · Abstract (English)
Mechanistic interpretation has greatly contributed to a more detailed understanding of generative language models, enabling significant progress in identifying structures that implement key behaviors through interactions between internal components. In contrast, interpretability in information retrieval (IR) remains relatively coarse-grained, and much is still unknown as to how IR models determine whether a document is relevant to a query. In this work, we address this gap by mechanistically analyzing how one commonly used model, a cross-encoder, estimates relevance. We find that the model extracts traditional relevance signals, such as term frequency and inverse document frequency, in early-to-middle layers. These concepts are then combined in later layers, similar to the well-known probabilistic ranking function, BM25. Overall, our analysis offers a more nuanced understanding of how IR models compute relevance. Isolating these components lays the groundwork for future interventions that could enhance transparency, mitigate safety risks, and improve scalability.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。