Understanding Matching Mechanisms in Cross-Encoders

TL;DR

This paper investigates the internal matching mechanisms of cross-encoder neural IR models, revealing how attention heads contribute to matching detection through causal analysis, thus providing interpretability insights.

Contribution

It introduces straightforward interpretability methods to analyze attention heads and elucidate the matching detection process in cross-encoders, moving beyond high-level explanations.

Findings

Attention heads play crucial roles in matching detection.

Causal insights reveal how attention influences predictions.

Interpretability methods clarify the matching process in neural IR models.

Abstract

Neural IR architectures, particularly cross-encoders, are highly effective models whose internal mechanisms are mostly unknown. Most works trying to explain their behavior focused on high-level processes (e.g., what in the input influences the prediction, does the model adhere to known IR axioms) but fall short of describing the matching process. Instead of Mechanistic Interpretability approaches which specifically aim at explaining the hidden mechanisms of neural models, we demonstrate that more straightforward methods can already provide valuable insights. In this paper, we first focus on the attention process and extract causal insights highlighting the crucial roles of some attention heads in this process. Second, we provide an interpretation of the mechanism underlying matching detection.