Reproducing and Extending Causal Insights Into Term Frequency Computation in Neural Rankers

TL;DR

This work investigates the causal underpinnings of relevance computation in neural rankers by reproducing and extending activation patching in information retrieval. It tests $TFC1$ and extends to $TFC2$ using diagnostic MS-MARCO datasets and a TAS-B ranker, locating term-frequency signals to specific attention heads (e.g., 0.9, 1.6, 2.3) and showing aggregation in the CLS token. Results show partial alignment with $TFC1$ and reveal that adherence to the axiom is inconsistent when term frequency is incrementally increased, with $TFC2$ behavior suggesting sublinear gains and dataset-sensitive effects. The findings highlight the sensitivity of causal conclusions to diagnostic design and motivate more robust, axiom-based interpretability work in neural information retrieval.

Abstract

Neural ranking models have shown outstanding performance across a variety of tasks, such as document retrieval, re-ranking, question answering and conversational retrieval. However, the inner decision process of these models remains largely unclear, especially as models increase in size. Most interpretability approaches, such as probing, focus on correlational insights rather than establishing causal relationships. The paper 'Axiomatic Causal Interventions for Reverse Engineering Relevance Computation in Neural Retrieval Models' by Chen et al. addresses this gap by introducing a framework for activation patching - a causal interpretability method - in the information retrieval domain, offering insights into how neural retrieval models compute document relevance. The study demonstrates that neural ranking models not only capture term-frequency information, but also that these representations can be localized to specific components of the model, such as individual attention heads or layers. This paper aims to reproduce the findings by Chen et al. and to further explore the presence of pre-defined retrieval axioms in neural IR models. We validate the main claims made by Chen et al., and extend the framework to include an additional term-frequency axiom, which states that the impact of increasing query term frequency on document ranking diminishes as the frequency becomes higher. We successfully identify a group of attention heads that encode this axiom and analyze their behavior to give insight into the inner decision-making process of neural ranking models.

Figures (6)

• Figure 1: Results of patching the residual stream, attention outputs, and MLP outputs for TFC1-I experiments with append and prepend perturbations. The injected token has the biggest effect on both appended and prepended query terms, while the information is aggregated in the CLS token in latest layers.
• Figure 2: Activation patching on individual attention heads for TFC1-I append and prepend experiments show that term-frequency information is related to attention heads 0.9, 1.6 and 2.3. These attention heads are active when patching the top 10% ranked documents (right), as opposed to patching the bottom 10% documents (left). A blue color denotes that patching a specific attention head recovers the perturbed performance, no color denotes that the patch recovers baseline performance, and red indicates that the patch recovers less than the baseline performance.
• Figure 3: The average attention scores for duplicate token heads show how the most relevant attention heads communicate to construct the relevance signal. More precisely, we see that information gets passed to the separator tokens, especially by node L2H3.
• Figure 4: Results of patching the residual stream, attention outputs and MLP outputs for TFC2-K1, TFC2-K3 and TFC2-K5 experiments. We observe that $t_{qterm-}$ has a significant impact for TFC2, while the injected token shows diminishing returns as the frequency of the injected term increases. As with TFC1, information seems to get aggregated in the latest layers under the CLS token. However, the impact is noticeable for a small K.
• Figure 5: Activation patching on individual attention heads for TFC2-K2, TFC2-K4, TFC2-K6, and TFC2-K8, comparing the top and bottom 10% ranked documents show that attention heads 1.0 and 1.9, as well as 1.6 and 0.9, are related to the TFC2 axiom.A blue color denotes that patching a specific attention head recovers the perturbed performance, no color denotes that the patch recovers baseline performance, and red indicates that the patch recovers less than the baseline performance.