On Survivorship Bias in MS MARCO

TL;DR

This paper investigates survivorship bias in MS MARCO, revealing that many unanswerable or discarded queries are valid, which distorts dataset distribution and evaluation scores, impacting model training and assessment.

Contribution

It identifies and analyzes survivorship bias in MS MARCO, demonstrating its effects on dataset quality, evaluation, and model performance, and provides insights for more complete annotation.

Findings

Discarded queries include many valid questions.

Survivorship bias distorts query distribution and evaluation scores.

Models trained on biased subsets perform worse on complete datasets.

Abstract

Survivorship bias is the tendency to concentrate on the positive outcomes of a selection process and overlook the results that generate negative outcomes. We observe that this bias could be present in the popular MS MARCO dataset, given that annotators could not find answers to 38--45% of the queries, leading to these queries being discarded in training and evaluation processes. Although we find that some discarded queries in MS MARCO are ill-defined or otherwise unanswerable, many are valid questions that could be answered had the collection been annotated more completely (around two thirds using modern ranking techniques). This survivability problem distorts the MS MARCO collection in several ways. We find that it affects the natural distribution of queries in terms of the type of information needed. When used for evaluation, we find that the bias likely yields a significant distortion of the absolute performance scores observed. Finally, given that MS MARCO is frequently used for model training, we train models based on subsets of MS MARCO that simulates more survivorship bias. We find that models trained in this setting are up to 9.9% worse when evaluated on versions of the dataset with more complete annotations, and up to 3.5% worse at zero-shot transfer. Our findings are complementary to other recent suggestions for further annotation of MS MARCO, but with a focus on discarded queries. Code and data for reproducing the results of this paper are available in an online appendix.