Interpretability Analysis of Domain Adapted Dense Retrievers
Goksenin Yuksel, Jaap Kamps
TL;DR
This paper tackles the problem of understanding how unsupervised domain adaptation affects dense retrievers by applying Integrated Gradients to generate instance-based and ranking-based explanations. Using FIQA and TREC-COVID, the authors show that domain-adapted models shift input attributions toward in-domain vocabulary and, in the case of TREC-COVID, place greater emphasis on document titles. The proposed IG-based methodology, including a novel baseline that exposes both query and document attributions, provides qualitative insights into how domain adaptation changes model behavior beyond traditional retrieval metrics. The work demonstrates the viability of IG for explaining dense retrievers and lays groundwork for more interpretable domain adaptation in neural IR, while acknowledging the need for more global attribution analysis and comparisons with other explanation methods.
Abstract
Dense retrievers have demonstrated significant potential for neural information retrieval; however, they exhibit a lack of robustness to domain shifts, thereby limiting their efficacy in zero-shot settings across diverse domains. Previous research has investigated unsupervised domain adaptation techniques to adapt dense retrievers to target domains. However, these studies have not focused on explainability analysis to understand how such adaptations alter the model's behavior. In this paper, we propose utilizing the integrated gradients framework to develop an interpretability method that provides both instance-based and ranking-based explanations for dense retrievers. To generate these explanations, we introduce a novel baseline that reveals both query and document attributions. This method is used to analyze the effects of domain adaptation on input attributions for query and document tokens across two datasets: the financial question answering dataset (FIQA) and the biomedical information retrieval dataset (TREC-COVID). Our visualizations reveal that domain-adapted models focus more on in-domain terminology compared to non-adapted models, exemplified by terms such as "hedge," "gold," "corona," and "disease." This research addresses how unsupervised domain adaptation techniques influence the behavior of dense retrievers when adapted to new domains. Additionally, we demonstrate that integrated gradients are a viable choice for explaining and analyzing the internal mechanisms of these opaque neural models.