Guardado en:
Detalles Bibliográficos
Autores principales: Lu, Meng, Chen, Catherine, Eickhoff, Carsten
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2502.04645
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866908669961043968
author Lu, Meng
Chen, Catherine
Eickhoff, Carsten
author_facet Lu, Meng
Chen, Catherine
Eickhoff, Carsten
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2502_04645
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Pathway to Relevance: How Cross-Encoders Implement a Semantic Variant of BM25
Lu, Meng
Chen, Catherine
Eickhoff, Carsten
Information Retrieval
Artificial Intelligence
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.
title Pathway to Relevance: How Cross-Encoders Implement a Semantic Variant of BM25
topic Information Retrieval
Artificial Intelligence
url https://arxiv.org/abs/2502.04645