Remining Hard Negatives for Generative Pseudo Labeled Domain Adaptation

David Rau; Goksenin Yuksel; Jaap Kamps

REVIEW 2 major objections 7 minor 24 references

Remining Hard Negatives for Generative Pseudo Labeled Domain Adaptation

T0 review · 2 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Periodically remining hard negatives with the in-training dense retriever, instead of keeping the initial static set, improves unsupervised domain adaptation of dense retrieval on most BEIR and LoTTE benchmarks.

desk verdict A straightforward, honest extension of GPL with modest but consistent gains; the false-negative worry is real but not fatal. read the letter →

arxiv 2501.14434 v1 pith:YKLMELQQ submitted 2025-01-24 cs.IR cs.LG

Goksenin Yuksel , David Rau , Jaap Kamps This is my paper · ORCID

classification cs.IRcs.LG

keywords denseretrievaldomainadaptationhardnegativemininggenerativepseudolabelingknowledgedistillationMarginMSEBEIRbenchmarkLoTTE

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that hard negatives for unsupervised domain adaptation of dense retrievers should be remined during training by the very model being adapted, rather than mined once at the start. The authors extend Generative Pseudo Labeling (GPL) into R-GPL, which refreshes the hard-negative index every 30,000 training steps, and report ranking gains on 13 of 14 BEIR datasets and 9 of 12 LoTTE test sets. The underlying insight is that a domain-adapted student retrieves documents that are more relevant to target queries, making them harder and more informative negatives. If correct, the result means domain adaptation can be improved without any extra labels, only by periodically refreshing the negative pool during distillation.

What carries the argument

The load-bearing mechanism is MarginMSE knowledge distillation combined with a dynamically refreshed hard-negative index. The loss trains the student to match the teacher's margin between the relevant document and the hard negative, $\mathrm{MSE}(DR(Q,D^+)-DR(Q,D^-),\, CE(Q,D^+)-CE(Q,D^-))$, and the re-mined negatives are top documents returned by the in-training student itself. Because those documents receive higher teacher relevance scores and sit closer to the query embedding, they create a harder margin to imitate, sharpening the training signal and broadening the set of document IDs seen in training.

What would settle it

On a dataset with human relevance judgments, audit the re-mined negative pool: if a substantial fraction of re-mined documents are judged relevant, and excluding those true positives from the negative pool eliminates R-GPL's gain over static GPL, the claimed mechanism would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that documents retrieved by a dense retriever after domain adaptation receive higher teacher relevance scores than those retrieved before adaptation, and that feeding these fresher, harder negatives back into MarginMSE distillation pushes the student further. R-GPL remines hard negatives every k steps using the in-training model, uses the cross-encoder teacher to score the new triples, and retrains the student on the refreshed margins. This yields consistent gains over static GPL on BEIR and LoTTE, with the largest gain on TREC-COVID (+4.4 NDCG@10). The paper also observes that the top-ranked documents returned after adaptation lie closer to the query embedding, so refreshing the index re-focuses training on the current decision boundary.

Load-bearing premise

That the documents the in-training model retrieves as hard negatives are truly non-relevant; if any of them are actually relevant, the MarginMSE loss is being optimized against a false label.

Editorial extensions

If this is right

R-GPL improves over static GPL on 13 of 14 BEIR datasets and 9 of 12 LoTTE test sets, with the largest BEIR gain on TREC-COVID (+4.4 NDCG@10).
Remining frequency is not critical: every tested interval (10k, 30k, 50k, 100k steps) improves over static GPL on TREC-COVID, FIQA, and SCIFACT, with 30k steps giving the best results.
Refreshing the negative index changes which document IDs appear in training, so the student sees a broader slice of the corpus rather than repeatedly optimizing against the same fixed set.
Each refresh briefly spikes the distillation loss, then the loss decays again, indicating the student is re-challenged by harder negatives and subsequently learns to imitate the teacher's margin on them.
The extra computation is confined to training; inference cost stays unchanged, which matters when dense retrieval is the first-stage ranker in a production pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The results suggest a positive feedback loop: better retrieval produces harder negatives, which in turn produce better retrieval; if the loop continues without label noise, unsupervised adaptation could keep improving over longer training horizons than static-GPL schedules allow.
The paper's analysis implies datasets where domain adaptation shifts the ranking most should benefit most from remining; this is a testable prediction that could be checked by correlating gain size with distribution shift in the student's top-10 results.
Because the paper does not verify that re-mined documents are truly irrelevant, a hidden risk is false negatives; an audit with human judgments could separate the gain from genuinely harder irrelevant documents versus accidentally relevant ones.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read R-GPL. The short version: it does what it says — periodically re-mining hard negatives with the model under adaptation yields small but consistent gains over GPL on most of BEIR and LoTTE. The idea isn't new; ANCE already did dynamic hard negative mining, and the paper credits it. What's useful here is the application to the GPL distillation setting, the LoTTE evaluation, and the analysis of score shifts. The score distribution plots and loss curves are informative and match the expected mechanics.

The central empirical claim holds up. The gains are modest — roughly +1.3 NDCG@10 on BEIR and +0.6 Success@5 on LoTTE — but they're consistent across 22 of 26 datasets, with query-level Wilcoxon tests. That's enough to take the method seriously.

The soft spots are real but not disqualifying. First, the remining frequency k=30k is tuned on TREC-COVID, FIQA, and SCIFACT, and those same datasets appear in the headline table. That's a mild selection bias. I'd want to see either all k results across all datasets or a proper holdout. Second, the stress-test on false negatives has some force. The paper's evidence that remined documents are 'harder' is entirely teacher scores from the same cross-encoder that creates the pseudo-labels. Some of those top documents could be actually relevant to the synthetic query — especially in datasets like NQ with multiple relevant passages — which would corrupt the MarginMSE target. The paper doesn't measure the false-negative rate. That said, this is a standard assumption in ANCE and DPR, and the method still improves even with some label noise, so I don't see it as a load-bearing flaw. It should be measured, but it doesn't invalidate the result. Third, there's no code and no multiple-seed variance. For a paper with gains this size, that's thin.

Overall, this is a competent, incremental empirical contribution. It deserves a serious referee. The reviewer should ask for the false-negative check, a cleaner treatment of the k selection, and either code or seed variance. The core finding — periodic remining helps GPL — is probably true and practically useful.

Referee Report

2 major / 7 minor

Summary. The paper proposes R-GPL, an extension of Generative Pseudo Labeling (GPL) for unsupervised domain adaptation of dense retrievers. Instead of keeping the initially mined hard negatives fixed throughout training, R-GPL periodically re-mines hard negatives using the student model that is itself undergoing domain adaptation. The authors evaluate on 14 BEIR datasets and 12 LoTTE test collections, reporting NDCG@10 gains over GPL in 13/14 BEIR datasets and Success@5 gains in 9/12 LoTTE datasets. They also present analyses of the teacher-predicted relevance scores of re-mined negatives, the effect of the re-mining frequency k, and the geometry of query/document embeddings before and after re-mining.

Significance. If the claimed effect is real, R-GPL is a simple and broadly applicable improvement over GPL: it changes only the hard-negative refresh schedule and requires no additional labels or model components, while consistently improving effectiveness on two external benchmarks. The paper deserves credit for evaluating on LoTTE in addition to BEIR, for reporting query-level significance tests on most datasets, and for analyzing the re-mining frequency. However, the central mechanistic claim—that re-mined documents are better hard negatives because they are harder yet non-relevant—is not verified against human judgments, and the hardness evidence uses the same teacher model that generates the pseudo-labels. In addition, the re-mining frequency k=30,000 is selected on three datasets that also appear in the headline results. These issues do not automatically invalidate the empirical gains, but they need to be addressed before the improvement can be confidently attributed to better hard-negative mining.

major comments (2)

[§4.1, §5, §6.2, Table 3] The paper treats re-mined documents D− as hard negatives, which §2.3 defines as irrelevant passages, but it never verifies that the re-mined documents are actually non-relevant. The evidence in §6.1 and Figures 2 and 6 is that the teacher cross-encoder assigns these documents higher relevance scores than the static miners' documents, and higher teacher scores are exactly what one would expect for documents that are truly relevant to the synthetic query. Because the same teacher supplies the pseudo-labels for the MarginMSE loss, this is a circular validation of 'hardness'. If a re-mined D− is actually relevant, the loss in §3.2 trains the student to push a relevant document away from the query, and the teacher margin CE(Q,D+)−CE(Q,D−) is no longer a valid distillation target. The authors should estimate the false-negative rate among re-mined negatives, for example by checking overlap with available qrels or by manually judging a sample, and should re-examine the mechanism claim in light of that estimate. The degradation on Arguana (Table 1) is at least consistent with label corruption and is not analyzed.
[§5, Tables 1–2] The re-mining frequency k=30,000 is selected in §6.2 using TREC-COVID, FIQA, and SCIFACT, which are three of the datasets that also appear in the headline results of Table 1. If the choice of k was made by inspecting test-set NDCG@10 on those datasets, then the reported gains for those three rows are optimistic and the broad '13/14 datasets' claim is weakened relative to a fully held-out hyperparameter choice. Moreover, Table 3 shows that for these three datasets every tested k improves over GPL, so the sensitivity to this choice is modest, but the authors should clarify whether k was selected on a validation split and, if not, either report results for all tested k values on the remaining datasets or explicitly exclude the tuning datasets from the headline count.

minor comments (7)

[§2.2] The benchmark name is written inconsistently as 'LoTTe', 'LoTTE', and 'LoTTE' in different places; please standardize to LoTTE.
[§6] There is a typo: 'syntetic queries' should be 'synthetic queries'.
[§7] In the opening sentence of §6, 'the the target domain' has a duplicated article.
[Table 3 caption] The sentence 'can of immense value' is missing a verb and should be 'can be of immense value'.
[§6.1] The notation 'GPL Xk' is not defined in the caption; please state explicitly that X is the number of training steps between re-mining operations.
[§6.1] The reference to Robertson (2007) and Arampatzis et al. (2009) is stated but not connected concretely to the observed histograms; the authors should explain how the normal/exponential score-distribution theory supports the conclusion that higher teacher scores imply better hard negatives rather than more relevant documents.
[§6.2] The y-axis labels in Figures 3 and 4 are not described in the captions; please add explicit axis titles or note the smoothing and units.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: R-GPL's main result is measured on external human-labeled benchmarks, and the remining mechanism is not derived from the teacher that supplies pseudo-labels.

full rationale

The central derivation chain is: GPL distills a cross-encoder teacher into a dense student using MarginMSE, and R-GPL refreshes the student's hard-negative index during training. The headline result (13/14 BEIR, 9/12 LoTTE) is evaluated with NDCG@10 and Success@5 on external, human-labeled benchmarks, so the evaluation is independent of the pseudo-labeling signal. The remining operation is defined by the in-training student's own retrieval, not by the teacher, and the MarginMSE loss is identical to GPL's; no parameter is fitted to the benchmark labels. The only tuning choice, remining frequency k, is selected on three test sets (Table 3), which is a test-set selection concern rather than circularity, and the paper reports gains for all k values (10k, 30k, 50k, 100k), so the main comparison is not forced by that choice. The hardness analysis in §6.1 (Fig. 2) indeed uses the same teacher cross-encoder that produces pseudo-labels, so it cannot by itself certify that remined documents are non-relevant; this is a validity caveat about false negatives, not a circular derivation of the benchmark result. The one self-citation (Arampatzis et al., 2009, with co-author Kamps) appears in a footnote about score distributions and is not load-bearing. No step reduces by construction to its own input.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The method introduces no new entities. Its novel content is a training-schedule modification to GPL, plus an empirical analysis. The main free parameter is the remining interval k, tuned on a subset of the test sets. The core assumptions are inherited from GPL's pseudo-labeling setup, plus the new assumption that the in-training student's top documents are valid hard negatives.

free parameters (1)

R-GPL remining frequency k = 30,000 steps
Selected based on NDCG@10 on TREC-COVID, FIQA, and SCIFACT (Table 3), then applied to all datasets. This is a hyperparameter tuned on a subset of the evaluation test sets.

assumptions (4)

domain assumption The teacher cross-encoder provides reliable relevance labels for generated query-document pairs in the target domain.
Inherited from GPL; R-GPL depends on these pseudo-labels for the MarginMSE training signal. If the teacher is miscalibrated on the target domain, both GPL and R-GPL fail.
domain assumption Top documents retrieved by the in-training student model are non-relevant and can serve as hard negatives.
R-GPL re-mines negatives from the student's own top-k retrieval; the paper does not manually verify that these are truly irrelevant, only that the teacher scores them as closer to relevant (harder). If the student retrieves actually relevant documents, they would be false negatives and corrupt training.
domain assumption Generated synthetic queries are relevant to the document used to generate them.
Inherited from GPL's query generation step; positive pairs are assumed valid. If the T5 query generator produces off-topic queries, the positive labels are wrong.
domain assumption The same remining frequency k generalizes across all target domains.
The paper fixes k=30,000 for all datasets based on 3 datasets, assuming the optimal interval does not vary wildly by corpus size or domain.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Remining Hard Negatives for Generative Pseudo Labeled Domain Adaptation." pith.science (2026). https://pith.science/paper/YKLMELQQ

@misc{pith2026250114434,
  author       = {Pith},
  title        = {Pith review of: Remining Hard Negatives for Generative Pseudo Labeled Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKLMELQQ}},
  note         = {Machine review of arXiv:2501.14434}
}

read the original 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. A state-of-the-art domain adaptation technique is Generative Pseudo Labeling (GPL). GPL uses synthetic query generation and initially mined hard negatives to distill knowledge from cross-encoder to dense retrievers in the target domain. In this paper, we analyze the documents retrieved by the domain-adapted model and discover that these are more relevant to the target queries than those of the non-domain-adapted model. We then propose refreshing the hard-negative index during the knowledge distillation phase to mine better hard negatives. Our remining R-GPL approach boosts ranking performance in 13/14 BEIR datasets and 9/12 LoTTe datasets. Our contributions are (i) analyzing hard negatives returned by domain-adapted and non-domain-adapted models and (ii) applying the GPL training with and without hard-negative re-mining in LoTTE and BEIR datasets.

Figures

Figures reproduced from arXiv: 2501.14434 by the authors.

**Figure 2.** Predicted relevancy scores for hard negatives retrieved by domain adapted model and hard negative miners. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Distillation training loss smoothed with an exponential moving average over length 50. Dashed lines [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Predicted cross encoder (CE) relevance margin between query (Q), relevant document [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 6.** Figure 6: Mean relevancy of returned hard negatives [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 5 canonical work pages

[1]

Avi Arampatzis and Stephen Robertson. 2011. https://doi.org/10.1007/S10791-010-9145-5 Modeling score distributions in information retrieval . Inf. Retr., 14(1):26--46

work page doi:10.1007/s10791-010-9145-5 2011
[2]

Avi Arampatzis, Stephen Robertson, and Jaap Kamps. 2009. https://doi.org/10.1007/978-3-642-04417-5\_13 Score distributions in information retrieval . In Advances in Information Retrieval Theory, Second International Conference on the Theory of Information Retrieval, ICTIR 2009, Cambridge, UK, September 10-12, 2009, Proceedings , volume 5766 of Lecture Not...

work page doi:10.1007/978-3-642-04417-5 2009
[3]

Krisztian Balog, Leif Azzopardi, Jaap Kamps, and Maarten de Rijke. 2006. https://doi.org/10.1007/978-3-540-74999-8\_101 Overview of webclef 2006 . In Evaluation of Multilingual and Multi-modal Information Retrieval, 7th Workshop of the Cross-Language Evaluation Forum, CLEF 2006, Alicante, Spain, September 20-22, 2006, Revised Selected Papers , volume 4730...

work page doi:10.1007/978-3-540-74999-8 2006
[4]

Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Nogueira. 2022. https://doi.org/10.1145/3477495.3531863 Inpars: Unsupervised dataset generation for information retrieval . In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 2387–2392, New York, NY, USA. Association ...

arXiv 2022
[5]

Xilun Chen, Kushal Lakhotia, Barlas Oğuz, Anchit Gupta, Patrick Lewis, Stan Peshterliev, Yashar Mehdad, Sonal Gupta, and Wen tau Yih. 2022. https://arxiv.org/abs/2110.06918 Salient phrase aware dense retrieval: Can a dense retriever imitate a sparse one? Preprint, arXiv:2110.06918

arXiv 2022
[6]

Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: retrieval-augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org

2020
[8]

Sebastian Hofstätter, Sophia Althammer, Michael Schröder, Mete Sertkan, and Allan Hanbury. 2021. https://arxiv.org/abs/2010.02666 Improving efficient neural ranking models with cross-architecture knowledge distillation . Preprint, arXiv:2010.02666

arXiv 2021
[9]

Johnson, M

J. Johnson, M. Douze, and H. Jegou. 2021. https://doi.org/10.1109/TBDATA.2019.2921572 Billion-scale similarity search with gpus . IEEE Transactions on Big Data, 7(03):535--547

arXiv 2021

Show all 24 references

[10]

Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.550 Dense passage retrieval for open-domain question answering . In Proceedings of the 2020 Conference on Empiric...

2020 doi
[11]

Omar Khattab and Matei Zaharia. 2020. https://doi.org/10.1145/3397271.3401075 Colbert: Efficient and effective passage search via contextualized late interaction over bert . In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Informatio...

2020
[12]

Ji Ma, Ivan Korotkov, Yinfei Yang, Keith Hall, and Ryan McDonald. 2021. https://doi.org/10.18653/v1/2021.eacl-main.92 Zero-shot neural passage retrieval via domain-targeted synthetic question generation . In Proceedings of the 16th Conference of the European Chapter of the Ass...

2021 doi
[13]

Craig Macdonald, Nicola Tonellotto, and Iadh Ounis. 2021. https://ceur-ws.org/Vol-2947/paper5.pdf On single and multiple representations in dense passage retrieval . In Proceedings of the 11th Italian Information Retrieval Workshop 2021, Bari, Italy, September 13-15, 2021, vol...

2021
[14]

Rodrigo Nogueira and Kyunghyun Cho. 2020. https://arxiv.org/abs/1901.04085 Passage re-ranking with bert . Preprint, arXiv:1901.04085

2020 arXiv
[15]

Ruiyang Ren, Yingqi Qu, Jing Liu, Xin Zhao, Qifei Wu, Yuchen Ding, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.1057 A thorough examination on zero-shot dense retrieval . In Findings of the Association for Computational Linguisti...

2023 doi
[16]

Stephen Robertson. 2007. https://doi.org/10.1007/978-3-540-71496-5\_7 On score distributions and relevance . In Advances in Information Retrieval, 29th European Conference on IR Research, ECIR 2007, Rome, Italy, April 2-5, 2007, Proceedings , volume 4425 of Lecture Notes in Co...

2007 doi
[17]

Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. 2022. https://doi.org/10.18653/v1/2022.naacl-main.272 C ol BERT v2: Effective and efficient retrieval via lightweight late interaction . In Proceedings of the 2022 Conference of the North Am...

2022 doi
[18]

Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. 2021 a . https://doi.org/10.18653/v1/2021.naacl-main.28 Augmented SBERT : Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks . In Proceedings of the 2021 Conference of th...

2021 doi
[19]

Nandan Thakur, Nils Reimers, Andreas R \" u ckl \' e , Abhishek Srivastava, and Iryna Gurevych. 2021 b . https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/65b9eea6e1cc6bb9f0cd2a47751a186f-Abstract-round2.html BEIR: A heterogeneous benchmark for zero-shot evalu...

2021
[20]

Kexin Wang, Nandan Thakur, Nils Reimers, and Iryna Gurevych. 2022. https://doi.org/10.18653/v1/2022.naacl-main.168 GPL : Generative pseudo labeling for unsupervised domain adaptation of dense retrieval . In Proceedings of the 2022 Conference of the North American Chapter of th...

2022 doi
[21]

Ji Xin, Chenyan Xiong, Ashwin Srinivasan, Ankita Sharma, Damien Jose, and Paul Bennett. 2022. https://doi.org/10.18653/v1/2022.findings-acl.316 Zero-shot dense retrieval with momentum adversarial domain invariant representations . In Findings of the Association for Computation...

2022 doi
[22]

Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. https://arxiv.org/abs/2007.00808 Approximate nearest neighbor negative contrastive learning for dense text retrieval . Preprint, arXiv:2007.00808

2020 arXiv
[23]

Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2023. https://doi.org/10.1145/3637870 Dense text retrieval based on pretrained language models: A survey . ACM Trans. Inf. Syst. Just Accepted

2023 doi
[24]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[25]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Avi Arampatzis and Stephen Robertson. 2011. https://doi.org/10.1007/S10791-010-9145-5 Modeling score distributions in information retrieval . Inf. Retr., 14(1):26--46

work page doi:10.1007/s10791-010-9145-5 2011

[2] [2]

Avi Arampatzis, Stephen Robertson, and Jaap Kamps. 2009. https://doi.org/10.1007/978-3-642-04417-5\_13 Score distributions in information retrieval . In Advances in Information Retrieval Theory, Second International Conference on the Theory of Information Retrieval, ICTIR 2009, Cambridge, UK, September 10-12, 2009, Proceedings , volume 5766 of Lecture Not...

work page doi:10.1007/978-3-642-04417-5 2009

[3] [3]

Krisztian Balog, Leif Azzopardi, Jaap Kamps, and Maarten de Rijke. 2006. https://doi.org/10.1007/978-3-540-74999-8\_101 Overview of webclef 2006 . In Evaluation of Multilingual and Multi-modal Information Retrieval, 7th Workshop of the Cross-Language Evaluation Forum, CLEF 2006, Alicante, Spain, September 20-22, 2006, Revised Selected Papers , volume 4730...

work page doi:10.1007/978-3-540-74999-8 2006

[4] [4]

Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Nogueira. 2022. https://doi.org/10.1145/3477495.3531863 Inpars: Unsupervised dataset generation for information retrieval . In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 2387–2392, New York, NY, USA. Association ...

arXiv 2022

[5] [5]

Xilun Chen, Kushal Lakhotia, Barlas Oğuz, Anchit Gupta, Patrick Lewis, Stan Peshterliev, Yashar Mehdad, Sonal Gupta, and Wen tau Yih. 2022. https://arxiv.org/abs/2110.06918 Salient phrase aware dense retrieval: Can a dense retriever imitate a sparse one? Preprint, arXiv:2110.06918

arXiv 2022

[6] [6]

Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: retrieval-augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org

2020

[7] [8]

Sebastian Hofstätter, Sophia Althammer, Michael Schröder, Mete Sertkan, and Allan Hanbury. 2021. https://arxiv.org/abs/2010.02666 Improving efficient neural ranking models with cross-architecture knowledge distillation . Preprint, arXiv:2010.02666

arXiv 2021

[8] [9]

Johnson, M

J. Johnson, M. Douze, and H. Jegou. 2021. https://doi.org/10.1109/TBDATA.2019.2921572 Billion-scale similarity search with gpus . IEEE Transactions on Big Data, 7(03):535--547

arXiv 2021