Research
Team Project + Independent ExtensionInformation Retrieval · Search · Model Evaluation

Lexical, Dense & Hybrid Information Retrieval

Keyword and meaning-based search found complementary evidence. Combining them improved retrieval on two public benchmarks; a fixed neural reranker then made the strongest system slower and worse.

Central findingHybrid retrieval improved aggregate MAP over the active lexical stage on both public benchmarks, while the tested cross-encoder reranker reduced quality and increased latency.

Evidence

The result in context

SciFact ΔMAP, hybrid vs active lexical stage
+0.0431
95% CI [0.0194, 0.0667].
NFCorpus ΔMAP, hybrid vs active lexical stage
+0.0331
95% CI [0.0260, 0.0406].
Estimated SciFact hybrid end-to-end latency
154.8ms
RRF total assembled from measured direct-system CPU timings.
Automated extension tests
22
On this page

Question

When does modern semantic retrieval genuinely improve on a strong lexical engine, and what does it cost in latency, complexity and interpretability?

A field-aware lexical search engine followed by an independent study of dense retrieval, rank fusion, neural reranking and latency on public benchmarks.

Part I · Team Project

A field-aware lexical search engine

The historical system built a positional index over licensed TREC Disk 4 & 5 / Robust04 coursework data, combining BM25F field weighting, phrase and proximity logic and query expansion across 249 topics.

The historical team was Blazej Olszta, Muhamad Husaam Ateeq, Max Monaghan and Sulaiman Bhatti. Its strongest BM25F + phrase/proximity configuration reached 0.1961 MAP, 0.4040 P@10 and 0.4033 nDCG@10.

Provenance

Historical data is not redistributed

The original evaluation was run locally using the licensed TREC coursework data and associated artefacts, which are not redistributed in the public repository. The historical phase is therefore not independently rebuildable from the public repository alone.

Part II · Independent Extension

Lexical and semantic retrieval were tested together

Plain BM25F scored 0.6374 MAP on SciFact and 0.1443 on NFCorpus. The active lexical ladder stage, BM25F + phrase/proximity, scored 0.6349 and 0.1444. Hybrid reciprocal-rank fusion reached 0.6780 and 0.1776.

The paired-bootstrap comparison uses the active lexical stage: SciFact ΔMAP +0.0431, 95% CI [0.0194, 0.0667]; NFCorpus ΔMAP +0.0331, 95% CI [0.0260, 0.0406].

Retrieval system explorer

Keyword and meaning are complementary

Compare aggregate quality, CPU latency and two licensed public SciFact query examples.

Benchmark
System

Keyword search

BM25F

MAP
0.6374
Measured mean CPU query latency
29.05ms
Strong field-aware lexical baseline.

Hybrid RRF vs active lexical stage

Benchmark query ID · 575

In domesticated populations of Saccharomyces cerevisiae, whole chromosome aneuploidy is very uncommon.
ΔAP +0.8571Semantic and lexical evidence are complementary for a long technical claim.

Direct-system CPU latency is measured. Hybrid RRF latency is estimated end to end; reranked latency is derived by adding measured reranker time. Examples are public SciFact claims licensed CC BY 4.0 and selected by absolute Hybrid RRF-minus-active-lexical ΔAP in the source query analysis. No document text is reproduced. AP is average precision for one query; MAP averages AP across benchmark queries.

Interactive evidence

Hybrid retrieval is evaluated against a controlled system ladder

MAP on SciFact and NFCorpus across six retrieval systems.

Read: Hybrid reciprocal-rank fusion reaches 0.6780 MAP on SciFact and 0.1776 on NFCorpus; gains differ materially by dataset.

Data table · 12 verified rows
Complete dataset
DatasetSystemMapMrrPrecision At10Ndcg At10System Dataset
ScifactBm250.6255890.6370260.0866670.66407Bm25 · Scifact
ScifactBm25f0.6373970.6495770.0873330.675244Bm25f · Scifact
ScifactBm25f Phrase Proximity0.6349010.6479880.0870.67285Bm25f Phrase Proximity · Scifact
ScifactDense0.6491670.6628770.0920.688526Dense · Scifact
ScifactHybrid Rrf0.6779630.6900020.0943330.71741Hybrid Rrf · Scifact
ScifactHybrid Rerank0.5722110.5786730.0846670.610178Hybrid Rerank · Scifact
NfcorpusBm250.1436260.5226070.2164090.308538Bm25 · Nfcorpus
NfcorpusBm25f0.1443080.5213840.2164090.309107Bm25f · Nfcorpus
NfcorpusBm25f Phrase Proximity0.1444250.5210030.215480.308816Bm25f Phrase Proximity · Nfcorpus
NfcorpusDense0.1642910.5261130.2414860.328166Dense · Nfcorpus
NfcorpusHybrid Rrf0.1775670.5607540.2495360.345822Hybrid Rrf · Nfcorpus
NfcorpusHybrid Rerank0.1686830.5298910.23870.329391Hybrid Rerank · Nfcorpus

Interactive evidence

The ablation ladder keeps each retrieval addition attributable

SciFact MAP in the explicit control order, from BM25 through neural reranking.

Read: Fusion improves the controlled lexical and dense baselines, while the neural reranker gives back quality in this frozen evaluation.

Data table · 6 verified rows
Complete dataset
DatasetSystemMapMrrPrecision At10Ndcg At10
ScifactBm250.6255890.6370260.0866670.66407
ScifactBm25f0.6373970.6495770.0873330.675244
ScifactBm25f Phrase Proximity0.6349010.6479880.0870.67285
ScifactDense0.6491670.6628770.0920.688526
ScifactHybrid Rrf0.6779630.6900020.0943330.71741
ScifactHybrid Rerank0.5722110.5786730.0846670.610178

Interactive evidence

Candidate depth reveals where later reranking can and cannot help

Recall at four candidate depths for lexical, dense and hybrid retrieval on both datasets.

Read: The diagnostic separates candidate generation from reranking: documents absent from the candidate set cannot be recovered downstream.

Data table · 24 verified rows
Representative preview rows; the full 24-row local dataset loads with the interactive chart.
DatasetSystemDepthRecall PctConfiguration
ScifactBm25f Phrase Proximity1079.011111Scifact · Bm25f Phrase Proximity · @10
ScifactBm25f Phrase Proximity5086.605556Scifact · Bm25f Phrase Proximity · @50
ScifactBm25f Phrase Proximity1,00096.5Scifact · Bm25f Phrase Proximity · @1000
ScifactDense5092.1Scifact · Dense · @50
ScifactHybrid Rrf1084.122222Scifact · Hybrid Rrf · @10
ScifactHybrid Rrf10096.833333Scifact · Hybrid Rrf · @100
NfcorpusBm25f Phrase Proximity5020.947963Nfcorpus · Bm25f Phrase Proximity · @50
NfcorpusBm25f Phrase Proximity1,00036.870095Nfcorpus · Bm25f Phrase Proximity · @1000
NfcorpusDense10029.948514Nfcorpus · Dense · @100
NfcorpusHybrid Rrf1016.95067Nfcorpus · Hybrid Rrf · @10
NfcorpusHybrid Rrf10030.749552Nfcorpus · Hybrid Rrf · @100
NfcorpusHybrid Rrf1,00061.373473Nfcorpus · Hybrid Rrf · @1000

Latency and ablation

The neural reranker was slower and worse

The tested cross-encoder reduced SciFact MAP from 0.6780 to 0.5722 and NFCorpus MAP from 0.1776 to 0.1687. Approximate CPU latency rose from 154.8ms to 553.3ms per SciFact query and from 21.2ms to 415.1ms on NFCorpus.

That negative result is useful: adding a more complex model is not automatically an upgrade.

Interactive evidence

Quality gains must be read alongside end-to-end latency

MAP with measured direct-system CPU latency, estimated Hybrid RRF totals and derived reranked totals.

Read: The neural reranker is slower and worse than hybrid RRF in the frozen study, so added model complexity is not treated as progress by default.

Direct systems are measured. Hybrid RRF is estimated end to end; Hybrid + reranker is derived by adding measured reranker time.

Data table · 12 verified rows
Complete dataset
DatasetSystemLatency MsLatency NatureMap
ScifactBm2519.408251Measured0.625589
ScifactBm25f29.045203Measured0.637397
ScifactBm25f Phrase Proximity141.595711Measured0.634901
ScifactDense10.671352Measured0.649167
ScifactHybrid Rrf154.826741Estimated0.677963
ScifactHybrid Rerank553.256321Derived0.572211
NfcorpusBm254.902424Measured0.143626
NfcorpusBm25f4.993294Measured0.144308
NfcorpusBm25f Phrase Proximity9.999905Measured0.144425
NfcorpusDense9.750033Measured0.164291
NfcorpusHybrid Rrf21.221645Estimated0.177567
NfcorpusHybrid Rerank415.052197Derived0.168683

Limitations

What this evidence does not establish

  • The historical TREC phase cannot be rebuilt from the public repository alone because licensed data and associated artefacts are not redistributed.
  • The extension conclusions are specific to SciFact, NFCorpus, the selected encoders and a fixed reranker configuration.

Source and reproducibility

Trace the evidence

Source code, evaluation outputs and supporting material are available in the repository.

View repository
  1. Public repositoryreports/figures/generated_summary.jsonCommit / evidence ID: 1cc2ea72b0d5fa8e94be28ef2c1e69626d0ad22b
  2. Historical coursework summaryreports/historical_coursework_summary.mdCommit / evidence ID: 1cc2ea72b0d5fa8e94be28ef2c1e69626d0ad22b
  3. Quality and latency frontierreports/figures/generated_summary.jsonCommit / evidence ID: 1cc2ea72b0d5fa8e94be28ef2c1e69626d0ad22b