[Paper Review] BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models

Nandan Thakur, Nils Reimers+3 more|arXiv (Cornell University)|Apr 17, 2021

Topic Modeling100 citations

TL;DR

BEIR presents a heterogeneous, zero-shot IR benchmark with 18 datasets across 9 tasks to evaluate 10 retrieval systems, revealing no single winner and highlighting generalization and efficiency trade-offs.

ABSTRACT

Existing neural information retrieval (IR) models have often been studied in homogeneous and narrow settings, which has considerably limited insights into their out-of-distribution (OOD) generalization capabilities. To address this, and to facilitate researchers to broadly evaluate the effectiveness of their models, we introduce Benchmarking-IR (BEIR), a robust and heterogeneous evaluation benchmark for information retrieval. We leverage a careful selection of 18 publicly available datasets from diverse text retrieval tasks and domains and evaluate 10 state-of-the-art retrieval systems including lexical, sparse, dense, late-interaction and re-ranking architectures on the BEIR benchmark. Our results show BM25 is a robust baseline and re-ranking and late-interaction-based models on average achieve the best zero-shot performances, however, at high computational costs. In contrast, dense and sparse-retrieval models are computationally more efficient but often underperform other approaches, highlighting the considerable room for improvement in their generalization capabilities. We hope this framework allows us to better evaluate and understand existing retrieval systems, and contributes to accelerating progress towards better robust and generalizable systems in the future. BEIR is publicly available at https://github.com/UKPLab/beir.

Motivation & Objective

• Motivate robust evaluation of IR models beyond in-domain tests across diverse domains and tasks.
• Provide a unified, zero-shot benchmark to study cross-domain generalization of lexical, sparse, dense, late-interaction, and re-ranking models.
• Enable fair comparison by standardizing data formats and evaluation metrics across tasks.
• Offer insights into which architectures generalize best and at what computational costs.

Proposed method

• Select 18 English zero-shot datasets from 9 retrieval tasks spanning diverse domains and data characteristics.
• Evaluate 10 state-of-the-art retrieval systems across five architectures: lexical, sparse, dense, late-interaction, and re-ranking.
• Use a standard data format (corpus, queries, qrels) and nDCG@10 as the primary evaluation metric.
• Benchmark implementations integrate with popular IR toolkits (Anserini, BM25, ColBERT, DPR, ANCE, TAS-B, GenQ, etc.).
• Analyze in-domain vs. zero-shot performance, efficiency (latency and index size), and annotation bias impacts.

Experimental results

Research questions

• RQ1 How well do diverse IR models generalize to out-of-distribution datasets across multiple tasks and domains?
• RQ2 Which architectures (lexical, sparse, dense, late-interaction, re-ranking) provide the strongest zero-shot performance on BEIR datasets?
• RQ3 What is the relationship between in-domain performance and cross-domain generalization, and how do efficiency considerations trade off with accuracy?

Key findings

• BM25 remains a strong zero-shot baseline across BEIR datasets.
• Re-ranking and late-interaction models often achieve the best zero-shot performance but incur high latency and memory costs.
• Dense and sparse retrieval models are faster but frequently underperform BM25 on many datasets.
• Cross-attentional re-ranking (BM25+CE) and ColBERT generally show better out-of-distribution generalization than base dense models.
• Domain adaptation and training losses influence zero-shot generalization, with TAS-B and GenQ showing strong results in some domains.
• Annotation bias affects evaluation, with lexical bias inflating hits for lexical methods and undervaluing non-lexical approaches; rectifying holes in TREC-COVID improves dense model scores.