MS-Shift: An Analysis of MS MARCO Distribution Shifts on Neural Retrieval

Simon Lupart, Thibault Formal, Stéphane Clinchant

Abstract

Pre-trained Language Models have recently emerged in Information Retrieval as providing the backbone of a new generation of neural systems that outperform traditional methods on a variety of tasks. However, it is still unclear to what extent such approaches generalize in zero-shot conditions. The recent BEIR benchmark provides partial answers to this question by comparing models on datasets and tasks that differ from the training conditions. We aim to address the same question by comparing models under more explicit distribution shifts. To this end, we build three query-based distribution shifts within MS MARCO (query-semantic, query-intent, query-length), which are used to evaluate the three main families of neural retrievers based on BERT: sparse, dense, and late-interaction -- as well as a monoBERT re-ranker. We further analyse the performance drops between the train and test query distributions. In particular, we experiment with two generalization indicators: the first one based on train/test query vocabulary overlap, and the second based on representations of a trained bi-encoder. Intuitively, those indicators verify that the further away the test set is from the train one, the worse the drop in performance. We also show that models respond differently to the shifts -- dense approaches being the most impacted. Overall, our study demonstrates that it is possible to design more controllable distribution shifts as a tool to better understand generalization of IR models. Finally, we release the MS MARCO query subsets, which provide an additional resource to benchmark zero-shot transfer in Information Retrieval.

Introduction

While Pre-trained Language Models (PLMs) like BERT have revolutionized Information Retrieval (IR), their ability to generalize to unseen scenarios remains a critical concern. Existing benchmarks like BEIR evaluate cross-dataset generalization but often fail to isolate the specific causes of performance degradation. This work introduces MS-Shift, a framework designed to analyze neural search models under controlled "explicit" distribution shifts—changes in query topics, intent, or length—within the MS MARCO dataset. By isolating these variables, we can pinpoint exactly where different architectures of neural retrievers lose effectiveness.

Methodology

The researchers partitioned the MS MARCO passage retrieval dataset (8.8M passages, 500k training queries) into subsets based on three specific attributes:

  1. Query Topics (Semantic): Five distinct topics identified via k-means clustering on DistilBERT embeddings (e.g., Medical vs. Public Figures).
  2. WH-Words (Intent): Grouping queries by starting words like "what" (definitions) or "how" (instructions).
  3. Query Length (Syntactic): Splitting queries into "short" and "long" based on median length.
t-SNE visualization of the five semantic query clusters identified in MS MARCO.

Evaluation followed a leave-one-out protocol: models were trained on all clusters except one, which served as the "unseen" test set. Four model families were compared: Dense Bi-encoders (single vector representations), Sparse Bi-encoders (e.g., SPLADE), Late Interaction (e.g., ColBERT), and Cross-encoders (e.g., monoBERT).

Key Insights and Results

The analysis reveals a clear hierarchy of robustness among neural architectures. Dense bi-encoders proved the most fragile, with performance dropping by up to 18% when encountering unseen topics. In contrast, sparse models like SPLADE demonstrated superior robustness, likely due to their reliance on lexical components that transfer better across similar vocabulary distributions.

Comparative performance (MRR@10) and relative loss across different neural retrieval models under semantic distribution shifts.

Key findings from the experimental results include:

  • Architecture Robustness: ColBERT and monoBERT consistently maintained the highest MRR@10 across shifts. As shown in the results table, the dense bi-encoder's relative loss (8.3% to 18.0%) significantly exceeded that of ColBERT (often below 4%).
  • Intent vs. Topic: Shifts in query intent (WH-words) caused more severe drops than semantic shifts. "How" queries, representing instrumental intent, were particularly challenging for zero-shot generalization, with SPLADE suffering a 26.8% relative loss when that intent was unseen.
  • Predictive Indicators: Vocabulary overlap (Jaccard Similarity) and embedding space distance (Model-based Similarity) correlate strongly with performance drops, offering a potential unsupervised method for predicting model reliability in new domains.

Conclusion

This study demonstrates that controllable distribution shifts are essential for understanding the generalization limits of IR models. While dense retrievers are highly sensitive to shifts, late-interaction and sparse models offer better zero-shot transfer. Future work will explore mitigation strategies like knowledge distillation and the use of unsupervised predictors to estimate system reliability before deployment in unlabelled environments.

Summary updated

Abstract

Pre-trained Language Models have recently emerged in Information Retrieval as providing the backbone of a new generation of neural systems that outperform traditional methods on a variety of tasks. However, it is still unclear to what extent such approaches generalize in zero-shot conditions. The recent BEIR benchmark provides partial answers to this question by comparing models on datasets and tasks that differ from the training conditions. We aim to address the same question by comparing models under more explicit distribution shifts. To this end, we build three query-based distribution shifts within MS MARCO (query-semantic, query-intent, query-length), which are used to evaluate the three main families of neural retrievers based on BERT: sparse, dense, and late-interaction – as well as a monoBERT re-ranker. We further analyse the performance drops between the train and test query distributions. In particular, we experiment with two generalization indicators: the first one based on train/test query vocabulary overlap, and the second based on representations of a trained bi-encoder. Intuitively, those indicators verify that the further away the test set is from the train one, the worse the drop in performance. We also show that models respond differently to the shifts – dense approaches being the most impacted. Overall, our study demonstrates that it is possible to design more controllable distribution shifts as a tool to better understand generalization of IR models. Finally, we release the MS MARCO query subsets, which provide an additional resource to benchmark zero-shot transfer in Information Retrieval.
BibTeX
@inproceedings{lupart2023ms,
  author = {Simon Lupart and Thibault Formal and Stéphane Clinchant},
  title = {{MS-Shift: An Analysis of MS MARCO Distribution Shifts on Neural Retrieval}},
  booktitle = {ECIR},
  year = {2023},
  doi = {10.1007/978-3-031-28244-7\_40},
}
Figure