Pith. sign in

REVIEW 3 major objections 6 minor 37 references

Beyond Quantile Methods: Improved Top-K Threshold Estimation for Traditional and Learned Sparse Indexes

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read By merging precomputed per-subset score prefixes and performing bounded index lookups, top-k threshold estimation lifts MUF from 0.917 to 0.986 on ClueWeb09B for k=10, and the same methods improve learned sparse indexes.

desk verdict Solid, incremental paper: the MUF gains are real, especially via lookups, and the learned-sparse evaluation is new, but the safety argument silently assumes non-negative term scores and QLD breaks that assumption as stated. read the letter →

arxiv 2412.10701 v1 pith:SYTFBG3V submitted 2024-12-14 cs.IR

Jinrui Gou , Yifan Liu , Minghao Shao , Torsten Suel This is my paper · ORCID
classification cs.IR
keywords top-kthresholdestimationquantilemethodslearnedsparseindexesMeanUnder-predictionFractionMaxScorequeryprocessinginvertedindexlookupsprefixmergingBM25
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 tries to establish that top-k threshold estimation — predicting the score of the k-th ranked document before running the query — can be made far more accurate than quantile-based methods by spending a little extra time and space on document-level information. Starting from precomputed per-term and per-subset quantile scores, the authors store sorted prefixes of the top documents for each selected term subset, then merge those prefixes while combining per-term scores and performing a bounded number of inverted-index lookups for missing scores. On ClueWeb09B with k=10, the mean under-prediction fraction (MUF) improves from 0.917 with the best quantile method to 0.986 with an access budget of 1k and 100 lookups; on queries with six or more terms it improves from 0.786 to 0.953–0.971. The same recipe works on learned sparse indexes built with DocT5Query expansion and DeepImpact, where score distributions differ strongly from BM25. If the claim holds, query processors such as MaxScore can start from a much tighter initial threshold and answer disjunctive top-k queries faster without risking overestimates.

What carries the argument

The load-bearing object is a budgeted prefix-merge estimator over precomputed subset prefixes. For frequent term subsets (singles, pairs, and some triples) the index stores a conjunctive top-scoring prefix: each entry is a docID plus the term scores for every term in the subset, sorted by total subset score. At query time, the estimator scans these prefixes up to an access budget ab, accumulates combined scores per docID, and then performs up to lb lookups into the inverted index to fill in missing term scores; the result is backed up by the best quantile lower bound. The merging step removes duplicate docIDs and the lookup step recovers contributions from terms whose subsets were not stored, which is what turns a coarse quantile bound into a near-exact threshold. The whole construction inherits the safety of quantile lower bounds as long as all term impact scores are non-negative.

What would settle it

Run Combine Scores with lookups on an index whose scoring function contains a negative term weight, compare each estimate against an exact top-k computation, and check whether any estimate exceeds the true k-th score; even one overestimate refutes the safety claim. For the standard non-negative scorers, instrument the algorithm and count overestimates over the TREC Terabyte query set; the paper's claim is zero overestimates for the non-sampling variants, so a single observed overestimate would also be a falsifier.

Watch

Extended reading notes

Core claim

The central discovery is that the accuracy lost by quantile methods is recoverable information: the identity of the documents behind the stored threshold scores and the way their term-level contributions add up. Quantile methods keep only one number per term subset; this paper keeps, for each selected subset, a sorted prefix of its highest-scoring documents together with all per-term scores, and turns threshold estimation into a budgeted merging problem. A query estimate is obtained by walking these prefixes in decreasing total score, accumulating per-document scores in a hash table, removing duplicate documents, and optionally looking up missing term scores in the inverted index; the k-th highest accumulated score is the answer. Because every stored subset is contained in the query, each partial sum is a lower bound on that document's full query score, so the estimate is safe — it never exceeds the true threshold, provided term scores are non-negative. The experiments show this closes most of the gap to the ideal MUF of 1.0, with the largest gains on long queries and on learned sparse indexes.

Load-bearing premise

The load-bearing premise is that all term impact scores are non-negative (and that every stored subset is contained in the query), so a sum over a subset of terms can never exceed a document's full query score; with negative term weights the accumulated estimate could overshoot the true threshold and the safe/no-overestimate property would collapse.

Editorial extensions

If this is right

  • For k=10 on ClueWeb09B with BM25, MUF improves from 0.917 (Q4_k) to 0.986 with ab=1k and lb=100, and MaxScore average query time drops from 7626 to 6833 microseconds; for k=100 the time drops from 10024 to 9155 microseconds.
  • On long queries with six or more terms, where quantile methods degrade most, the k=10 MUF jumps from 0.786 to 0.953 with ab=200 and to 0.971 with ab=500.
  • For learned sparse indexes, DeepImpact k=10 MUF improves from 0.885 to 0.952 and MaxScore time drops from 3668 to 3415 microseconds, so the approach transfers beyond BM25-style scores.
  • For large k (1000), small and medium prefix structures give little benefit alone, but adding sampling at s=0.02 or s=0.05 with a 0.01% overestimate target restores substantial gains.
  • Lookups matter more than duplicate removal or score combination: the biggest single jump in accuracy comes from performing inverted-index lookups, and lb at half of ab still gives decent results.

Reading between the lines

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

  • The paper leaves implicit that the method effectively converts threshold estimation into a budgeted top-k execution over compressed per-subset lists, so any future improvement in accumulator-based top-k merging should transfer directly to estimation accuracy at fixed budgets.
  • Because pairs dominate the gains and triples/quadruples add little, a plausible extension is adaptive subset selection: choose prefix types based on query-length distribution and available space rather than all co-occurring subsets from a log.
  • The success on DeepImpact suggests the technique may extend to other learned sparse models with query-dependent weights, such as SPLADE or uniCOIL, once their query-time weights are decomposed into per-term impacts; the paper notes they require additional ideas but does not test them.
  • One could test a per-query adaptive lookup budget that allocates lb according to query length or number of available prefixes, since the paper shows long queries benefit most and lookup cost per estimate is only 20–40 nanoseconds.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper studies top-k threshold estimation, i.e., estimating the score of the k-th highest document for a disjunctive query, and proposes three enhancements over quantile-based baselines: Remove Duplicates, Combine Scores, and Adding Lookups, plus an optional random-sampling wrapper. The methods store prefixes of high-scoring documents for query-term subsets and combine scores with accumulator structures and inverted-index lookups. The evaluation on ClueWeb09B and MSMARCO under BM25, QLD, DocT5Query, and DeepImpact reports MUF improvements (e.g., from 0.917 to 0.986 for k=10 on ClueWeb09B, and from 0.786 to 0.953-0.971 for long queries), reduced space for practical configurations, and faster MaxScore query processing when the estimates are used as initial thresholds.

Significance. If the safety guarantee holds, the paper offers a practical advance in threshold estimation, with a clear algorithmic idea and an unusually broad experimental matrix spanning two datasets and four ranking functions, including learned sparse indexes. The lower-bound arguments for Remove Duplicates and Combine Scores are clearly reasoned under the standard non-negativity assumption, and the paper is the first to evaluate threshold estimation on DocT5Query and DeepImpact indexes, which is a useful contribution for the learned-sparse-retrieval community. The main gaps are that the non-negativity hypothesis underlying the safety proof is never stated, and the evaluation omits a standalone sampling baseline and observed overestimate rates for sampling variants; both are needed to fully support the paper's central claims.

major comments (3)
  1. [Section III (Combine Scores and Adding Lookups), Figures 6 and 9] The zero-overestimate guarantee for the prefix-based methods requires the implicit assumption that every per-term impact sc(t,d) is non-negative. This assumption is not stated in Section II or III, and it fails for the standard Dirichlet query-likelihood scorer sc(q,d)=sum_t log((tf_{t,d}+mu P(t|C))/(|d|+mu)), where each term contribution is negative for typical parameters. With negative term impacts, a partial sum over a subset of query terms can exceed the full query score, so the accumulated estimate can overestimate the true threshold. Since QLD is one of the four rankers in Figures 6 and 9, and MUF is by definition averaged only over queries that do not overestimate, the QLD results are not interpretable as reported. The authors must either state the non-negativity assumption and verify it for each of BM25, QLD, DocT5Query, and DeepImpact, or document the exact positive reparameterization of QLD used in their implementation, and they should report observed overestimation rates for QLD.
  2. [Section IV (comparison protocol)] The experiments compare only against the Q4_k quantile baseline, not against a standalone random-sampling estimator, even though Section II presents sampling from [11] as a competitive state-of-the-art method and the abstract explicitly positions the work relative to sampling-based methods. Because the claimed contribution is to go beyond quantile methods against the state of the art, the omission of a sampling baseline leaves the relative claim unsupported. At minimum, the sampling-integrated variant should be compared with plain sampling at the same target overestimation probability, and the observed overestimate rate should be reported for both.
  3. [Section IV (Figures 5, 8, and 9)] When the sampling wrapper is used, MUF values are reported without any table of observed overestimation rates. MUF is defined as an average over queries that do not overestimate, so a high MUF can coexist with a nontrivial overestimate rate; the paper currently only states the a priori target probability of 0.01% used to choose k'. Reporting the observed overestimation rate for each sampling configuration is necessary to judge the accuracy/safety trade-off and to compare fairly with the zero-overestimate prefix-based methods.
minor comments (6)
  1. [Section II] The acronym MUF is defined both as 'Mean Under prediction Fraction' and later as 'Mean Underestimation Rate'; please use one consistent definition and expansion.
  2. [Table I] The table header contains 'Clueweb09a'; this should be 'ClueWeb09B'.
  3. [Section IV-B] The statement that 'precise definitions of these configurations will be made available as part of the dataset release' gives no artifact identifier or URL; please include the configuration parameters in the paper or a supplement for reproducibility.
  4. [Section IV (setup and Tables II-III)] There are small typos: 'ALLookups' should be 'All lookups', and the table captions say 'in nanosecond' where 'nanoseconds' is intended.
  5. [Section IV (Table IV)] The text says the BM25 ClueWeb09B speedup is 'more than 800 us', but the table row shows 7626-6833=793 us; please correct the rounding or the text.
  6. [Figures 1-9] Adding error bars or confidence intervals for the MUF values would help assess the stability of the reported differences, particularly in Figure 5 where some curves are close.

Circularity Check

1 steps flagged · score 2.0 of 10

One explicit by-definition self-backup; central MUF gains are empirically independent.

  1. self definitional [Section III, Adding Lookups (Discussion paragraph)]
    "We also note that our implementations, unless stated otherwise, are backed up by a state-of-the-art quantile method. That is, the estimates we return are the maximum of the estimate returned by our method and the quantile method. Because our method must store larger prefixes, it cannot cover as many subsets as a quantile method that needs to only store a single score per prefix. Thus, for some queries, the quantile method can outperform our method without backup. Consequently, our methods are by definition as least as precise as the quantile method."

    The property 'at least as precise as the quantile method' is not derived from the method's design; it is true by construction because the final estimate is defined as the maximum of the proposed estimate and the Q4_k baseline. This is a self-definitional statement explicitly labeled 'by definition' in the paper. It is minor and transparent, and it is not the paper's central claim: the reported MUF improvements are measured empirically on ClueWeb09B and MSMARCO rather than being algebraically forced by the max-backup construction.

full rationale

The paper's central contribution is an experimental study of threshold-estimation extensions over quantile baselines. The baseline Q4_k is from a paper co-authored by one of the current authors, but it is implemented and tested directly, and the proposed methods' MUF gains are not forced by the baseline's definition. The only self-definitional element is the explicit max-backup statement, which the paper itself admits is 'by definition' and which does not affect the empirical comparison. The unstated non-negativity assumption for the safety guarantee of Combine Scores and Lookups is a correctness risk, especially for QLD, but it is not circularity: even if that guarantee fails, the method's outputs are not equal to its inputs by construction. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors, and no known result is merely renamed. The circularity score is therefore 2, reflecting the minor, non-load-bearing self-citation and the acknowledged by-definition backup, while the core MUF results retain independent experimental content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; the methods reuse standard inverted index postings and term scores. The free parameters are tuning knobs, not fitted constants, and the axioms are standard domain assumptions for IR systems.

free parameters (5)
  • access budget ab = 100 to 5000 (0 for baseline)
    Number of postings fetched from prefixes per query; chosen by experimenters to explore time/accuracy tradeoff; the central MUF results depend on ab.
  • lookup budget lb = 0 to ab (often lb=ab for full lookups)
    Number of top hash-table entries receiving index lookups; a tuning knob; benefits require lb > k.
  • prefix depth per subset type = 10k for singles/pairs, 4k triplets, 3k quadruplets in Part I; smaller adaptive depths in Part II
    Memory vs accuracy tradeoff; not derived from theory.
  • sampling rate s = 0.02 and 0.05 in experiments
    Used in Adding Sampling variant to reduce k to k' on a sample; chosen to keep overestimation probability at 0.01%.
  • effective top-k on sample k' = Derived from s and overestimation bound (not explicitly listed)
    Selected to limit overestimation probability to 0.01%; a free design choice.
assumptions (4)
  • domain assumption Additive scoring: sc(q,d) = sum_{t in q} sc(t,d)
    Stated in Section II Problem Definition; required by all methods.
  • domain assumption Term scores are non-negative
    Needed for subset score sums to lower-bound full query score (Combine Scores and Lookups); not stated explicitly; holds for BM25, QLD, DeepImpact.
  • domain assumption Document sampling is independent and uniform at rate s
    Used to derive overestimation probability O in Section III (Adding Sampling).
  • domain assumption Training query log (AOL) represents test queries
    Determines which subsets get prefixes; test queries are from TREC 2005; coverage of unseen terms not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Quantile Methods: Improved Top-K Threshold Estimation for Traditional and Learned Sparse Indexes." pith.science (2026). https://pith.science/paper/SYTFBG3V

@misc{pith2026241210701,
  author       = {Pith},
  title        = {Pith review of: Beyond Quantile Methods: Improved Top-K Threshold Estimation for Traditional and Learned Sparse Indexes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SYTFBG3V}},
  note         = {Machine review of arXiv:2412.10701}
}
read the original abstract

Top-k threshold estimation is the problem of estimating the score of the k-th highest ranking result of a search query. A good estimate can be used to speed up many common top-k query processing algorithms, and thus a number of researchers have recently studied the problem. Among the various approaches that have been proposed, quantile methods appear to give the best estimates overall at modest computational costs, followed by sampling-based methods in certain cases. In this paper, we make two main contributions. First, we study how to get even better estimates than the state of the art. Starting from quantile-based methods, we propose a series of enhancements that give improved estimates in terms of the commonly used mean under-prediction fraction (MUF). Second, we study the threshold estimation problem on recently proposed learned sparse index structures, showing that our methods also work well for these cases. Our best methods substantially narrow the gap between the state of the art and the ideal MUF of 1.0, at some additional cost in time and space.

Figures

Figures reproduced from arXiv: 2412.10701 by the authors.

Figure 1
Figure 1. Comparison of methods for k = 10 and k = 1000, for ClueWeb09B. We first discuss results for k = 10, shown in solid lines. We see at the bottom of the chart that Remove Duplicates provides basically no benefit over Q4 k for this case as the curves are on top of each other. This is not surprising as in many cases the top 10 postings are chosen from the same prefix, resulting in the same estimate as Q4 k , while Remove… view at source ↗
Figure 3
Figure 3. Comparing singles, pairs, triples, and quadruples for MSMARCO. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. looks at the lookup budget lb as a fraction of the access budget ab, for k = 10 on ClueWeb09B. Of course, lb = ab performs best, but even 50% gives decent results [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (5 more)
Figure 2
Figure 2. Figure 2: Comparing singles, pairs, triples, and quadruples for ClueWeb09B. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png]
Figure 5
Figure 5. Figure 5: MUF of different prefix configurations for [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparison of singles, pairs, triples, and quadruples for [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 9
Figure 9. Figure 9: MUF of different prefix configurations for [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 8
Figure 8. Figure 8: MUF of different prefix configurations for [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 35 canonical work pages

  1. [20]

    Fast first-phase candidate generation for cascading rankers,

    Q. Wang, C. Dimopoulos, and T. Suel, “Fast first-phase candidate generation for cascading rankers,” in Proc. SIGIR, 2016

    work page 2016

  2. [11]

    A comparison of top- k threshold estimation techniques for disjunctive query processing,

    A. Mallia, M. Siedlaczek, M. Sun, and T. Suel, “A comparison of top- k threshold estimation techniques for disjunctive query processing,” in Proc. CIKM, 2020

    work page 2020

  3. [1]

    A cascade ranking model for efficient ranked retrieval,

    L. Wang, J. Lin, and D. Metzler, “A cascade ranking model for efficient ranked retrieval,” in Proc. SIGIR, 2011

    work page 2011

  4. [2]

    Query evaluation: strategies and optimizations,

    H. Turtle and J. Flood, “Query evaluation: strategies and optimizations,” Information Processing & Management , vol. 31, no. 6, 1995

    work page 1995

  5. [3]

    Efficient query evaluation using a two-level retrieval process,

    A. Z. Broder, D. Carmel, M. Herscovici, A. Soffer, and J. Zien, “Efficient query evaluation using a two-level retrieval process,” in Proc. CIKM , 2003

    work page 2003

  6. [4]

    Faster top-k document retrieval using block-max indexes,

    S. Ding and T. Suel, “Faster top-k document retrieval using block-max indexes,” in Proc. SIGIR, 2011

    work page 2011

  7. [5]

    Optimizing top-k document retrieval strategies for block-max indexes,

    C. Dimopoulos, S. Nepomnyachiy, and T. Suel, “Optimizing top-k document retrieval strategies for block-max indexes,” in Proc. WSDM, 2013

    work page 2013

  8. [6]

    Heuristics to improve the BMW method and its variants,

    L. L. S. de Carvalho, E. S. de Moura, C. M. Daoud, and A. S. da Silva, “Heuristics to improve the BMW method and its variants,” JIDM, vol. 6, no. 3, 2015

    work page 2015

Show all 37 references
  1. [7]

    Faster blockmax wand with variable-sized blocks,

    A. Mallia, G. Ottaviano, E. Porciani, N. Tonellotto, and R. Venturini, “Faster blockmax wand with variable-sized blocks,” in Proc. SIGIR , 2017

    2017

  2. [8]

    Accelerated query processing via similarity score prediction,

    M. Petri, A. Moffat, J. Mackenzie, J. S. Culpepper, and D. Beck, “Accelerated query processing via similarity score prediction,” in Proc. SIGIR, 2019

    2019

  3. [9]

    Split-lists and initial thresholds for wand- based search,

    A. Kane and F. W. Tompa, “Split-lists and initial thresholds for wand- based search,” in Proc. SIGIR, 2018

    2018

  4. [10]

    A candidate filtering mechanism for fast top-k query processing on modern cpus,

    C. Dimopoulos, S. Nepomnyachiy, and T. Suel, “A candidate filtering mechanism for fast top-k query processing on modern cpus,” in Proc. SIGIR, 2013

    2013

  5. [12]

    Selective search: Efficient and effective search of large textual collections,

    A. Kulkarni and J. Callan, “Selective search: Efficient and effective search of large textual collections,” ACM TOIS, vol. 33, no. 4, 2015

    2015

  6. [13]

    Multi-tier architecture for web search engines,

    K. M. Risvik, Y . Aasheim, and M. Lidal, “Multi-tier architecture for web search engines,” in Proc. 1st Latin Am. Web Congress , 2003

    2003

  7. [14]

    Optimal web-scale tiering as a flow problem,

    G. Leung, N. Quadrianto, A. J. Smola, and K. Tsioutsiouliklis, “Optimal web-scale tiering as a flow problem,” in Proc. NIPS, 2010

    2010

  8. [15]

    Efficiency trade-offs in two-tier web search systems,

    R. Baeza-Yates, V . Murdock, and C. Hauff, “Efficiency trade-offs in two-tier web search systems,” in Proc. SIGIR, 2009

    2009

  9. [16]

    Taily: Shard selection using the tail of score distributions,

    R. Aly, D. Hiemstra, and T. Demeester, “Taily: Shard selection using the tail of score distributions,” in Proc. SIGIR, 2013

    2013

  10. [17]

    Caching scores for faster query process- ing with dynamic pruning in search engines,

    E. Yafay and I. S. Altingovde, “Caching scores for faster query process- ing with dynamic pruning in search engines,” in Proc. CIKM, 2019

    2019

  11. [18]

    Sushi: Scoring scaled samples for server selection,

    P. Thomas and M. Shokouhi, “Sushi: Scoring scaled samples for server selection,” in Proc. SIGIR, 2009

    2009

  12. [19]

    Optimal aggregation algorithms for middleware,

    R. Fagin, A. Lotem, and M. Naor, “Optimal aggregation algorithms for middleware,” in Proc. SIGMOD, 2001

    2001

  13. [21]

    From doc2query to doctttttquery,

    R. Nogueira and J. Lin, “From doc2query to doctttttquery,” Online preprint, 2019

    2019

  14. [22]

    Learning passage impacts for inverted indexes,

    A. Mallia, O. Khattab, T. Suel, and N. Tonellotto, “Learning passage impacts for inverted indexes,” in Proc. SIGIR, 2021

    2021

  15. [23]

    Shard ranking and cutoff estimation for topically partitioned collections,

    A. Kulkarni, A. S. Tigelaar, D. Hiemstra, and J. Callan, “Shard ranking and cutoff estimation for topically partitioned collections,” in Proc. CIKM, 2012

    2012

  16. [24]

    Using conjunctions for faster disjunctive top-k queries,

    M. Siedlaczek, A. Mallia, and T. Suel, “Using conjunctions for faster disjunctive top-k queries,” in Proc. WSDM, 2022

    2022

  17. [25]

    Exploring the magic of wand,

    M. Petri, J. S. Culpepper, and A. Moffat, “Exploring the magic of wand,” in Proc. ADCS, 2013

    2013

  18. [26]

    Fast disjunctive candidate generation using live block filtering,

    A. Mallia, M. Siedlaczek, and T. Suel, “Fast disjunctive candidate generation using live block filtering,” in Proc. WSDM, 2021

    2021

  19. [27]

    COIL: revisit exact lexical match in information retrieval with contextualized inverted list,

    L. Gao, Z. Dai, and J. Callan, “COIL: revisit exact lexical match in information retrieval with contextualized inverted list,” in Proc. NAACL- HLT, 2021

    2021

  20. [28]

    A few brief notes on DeepImpact, COIL, and a conceptual framework for information retrieval techniques,

    J. Lin and X. Ma, “A few brief notes on DeepImpact, COIL, and a conceptual framework for information retrieval techniques,” Preprint: arXiv:2106.14807, 2021

    2021 arXiv

  21. [29]

    SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking,

    T. Formal, B. Piwowarski, and S. Clinchant, “SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking,” in Proc. SIGIR, 2021

    2021

  22. [30]

    Wacky weights in learned sparse representations and the revenge of score-at-a-time query evaluation,

    J. Mackenzie, A. Trotman, and J. Lin, “Wacky weights in learned sparse representations and the revenge of score-at-a-time query evaluation,” preprint arXiv:2110.11540, 2021

    2021 arXiv

  23. [31]

    Faster learned sparse retrieval with guided traversal,

    A. Mallia, J. Mackenzie, T. Suel, and N. Tonellotto, “Faster learned sparse retrieval with guided traversal,” in Proc. SIGIR, 2022

    2022

  24. [32]

    Ioqp: A simple impact- ordered query processor written in rust,

    J. Mackenzie, M. Petri, and L. Gallagher, “Ioqp: A simple impact- ordered query processor written in rust,” in Proc. DESIRES, 2022

    2022

  25. [33]

    Accelerating learned sparse indexes via term impact decomposition,

    J. Mackenzie, A. Mallia, A. Moffat, and M. Petri, “Accelerating learned sparse indexes via term impact decomposition,” in Proc. EMNLP, 2022

    2022

  26. [34]

    PISA: performant indexes and search for academia,

    A. Mallia, M. Siedlaczek, J. Mackenzie, and T. Suel, “PISA: performant indexes and search for academia,” in Proc. OSIRRC@SIGIR, 2019

    2019

  27. [35]

    Quasi-succinct indices,

    S. Vigna, “Quasi-succinct indices,” in Proc. WSDM, 2013

    2013

  28. [36]

    Ms marco: A human generated machine reading compre- hension dataset,

    T. Nguyen, M. Rosenberg, X. Song, J. Gao, S. Tiwary, R. Majumder, and L. Deng, “Ms marco: A human generated machine reading compre- hension dataset,” choice, vol. 2640, 2016

    2016

  29. [37]

    A language modeling approach to information retrieval,

    J. M. Ponte and W. B. Croft, “A language modeling approach to information retrieval,” in Proc. SIGIR, 1998

    1998

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.