Home / Curated / Architectures / uniCOIL Unified Contextualized Inverted List Sparse Retrieval

uniCOIL Unified Contextualized Inverted List Sparse Retrieval

Architectures AIMech:uniCOILUnifiedContextualizedInvertedList DRAFT

Aliases: uniCOIL, Unified Contextualized Inverted List, uniCOIL Sparse Impact Retrieval

A learned sparse retrieval architecture that extends Contextualized Inverted List (COIL) by reducing the token representation dimension to d = 1 so the Transformer encoder outputs scalar contextualized term weights, applying a ReLU operation on the output term weights to guarantee non-negativity, expanding corpus documents prior to encoding with doc2query-T5 predicted queries to resolve exact-match vocabulary mismatch, and quantizing both document and query term weights into 8-bit integer impact scores that are indexed and evaluated via inner products in unmodified standard inverted indexes such as Lucene/Anserini.

Identity Notes & Concept Scope Distinct from COIL (AIMech:COILContextualizedInvertedList), which stores d = 32 real-valued contextualized token vectors requiring custom vector-inverted-list infrastructure (~40 GB index on MS MARCO) and allows negative weights; distinct from DeepCT (AIMech:DeepCTContextAwareTermImportanceEstimationForFirstStageRetrieval), which trains a regression model on pseudo-ground-truth term frequencies without document expansion or query-term weighting and still relies on BM25's IDF/length-normalization formula; and distinct from SPLADE (AIMech:SPLADESparseLexicalAndExpansionModel), which performs end-to-end MLM-logit vocabulary expansion rather than combining d = 1 COIL scalar impact weighting with external doc2query-T5 expansion.

Claims & Grounded Evidence

4 Claims

Mechanism #unicoil-scalar-relu-projection-and-8bit-impact-quantization

uniCOIL reduces COIL's token representation dimension from d = 32 to d = 1 to produce scalar term weights, adds a ReLU activation on the output term weights to force non-negative values amenable to inverted-index traversal, and quantizes both document and query term weights into 8-bit integer impact scores (Section 3), compressing the MS MARCO passage index to 1.3 GB in standard Lucene/Anserini compared to ~40 GB for COIL-tok, 26 GB for TCT-ColBERTv2, and 154 GB for ColBERT (Section 4).

supports Source: A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques · Locator: Section 3 & Section 4, pp. 3-5
supports Source: A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques · Locator: unicoil_term_impacts() and inverted_index_impact_score() verification of ReLU clamping and 8-bit quantization

Mechanism #unicoil-doc2query-t5-expansion-compensating-scalar-degeneration

Within the two-component sparse learned retrieval decomposition (expansion model + term weighting model, Section 2), unexpanded uniCOIL achieves 0.315 MRR@10 on MS MARCO Dev (below 32-dimensional COIL-tok's 0.341), but combining uniCOIL with doc2query-T5 document expansion raises MRR@10 to 0.352—surpassing both unexpanded COIL-tok (0.341) and DeepImpact (0.326) without requiring multi-dimensional weight vectors or custom retrieval infrastructure (Section 3, Table 2).

supports Source: A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques · Locator: Sections 2-3, Table 2 rows (2a)-(2h), pp. 2-4
supports Source: A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques · Locator: main() verification that doc2query-expanded uniCOIL outperforms unexpanded uniCOIL under vocabulary mismatch

Application #unicoil-tct-colbertv2-dense-sparse-hybrid-fusion-application

When top-1000 sparse candidates from doc2query-T5-expanded uniCOIL are fused with top-1000 dense candidates from TCT-ColBERTv2 by normalizing both score distributions into range(0, 1) and computing s_dense + alpha * s_sparse, the hybrid system achieves 0.378 MRR@10 on MS MARCO Passage Dev—matching TCT-ColBERTv2 + DeepImpact (0.378) and outperforming standalone RocketQA (0.370), ColBERT (0.360), and COIL-full (0.355) (Section 3, Table 2 rows (3a)-(4g)).

supports Source: A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques · Locator: Section 3, Table 2 rows (3a)-(4g), pp. 3-4
supports Source: A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques · Locator: minmax_normalize() and hybrid fusion verification

Limitation #unicoil-external-expansion-dependency-and-exact-match-limitation

Because uniCOIL's scalar token weighting only assigns impact scores to tokens present in its input text, without an external document-expansion stage (such as doc2query-T5) it remains restricted to exact lexical matching and suffers a 2.6-point MRR@10 drop relative to 32-dimensional COIL-tok (0.315 vs. 0.341, Section 2 & Section 3, Table 2).

supports Source: A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques · Locator: Section 2 (Expansions for sparse representation) & Section 3, Table 2 rows (2e) & (2g), pp. 2-4

⚠️ Curated Limitations & Negative Results

Without external document expansion (such as doc2query-T5), uniCOIL cannot score terms absent from the document and underperforms 32-dimensional COIL-tok (0.315 vs. 0.341 MRR@10 on MS MARCO Dev).
Relies on a two-step offline indexing pipeline (sequence-to-sequence document expansion followed by Transformer impact encoding and quantization) rather than single-model end-to-end vocabulary expansion.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	related_match	uniCOIL uses a pretrained Transformer neural network encoder with a 1-dimensional ReLU projection head to predict contextualized scalar term impacts.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r775-primary-pdf`	A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques Jimmy Lin, Xueguang Ma	primary paper full_text	ID: https://arxiv.org/bibtex/2106.14807 sha256:c5189bebaab0…
`r775-arxiv-abs`	A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques (arXiv abstract) Jimmy Lin, Xueguang Ma	primary paper abstract_only	ID: https://arxiv.org/abs/2106.14807 sha256:d4f96c941552…
`r775-metadata-cr`	A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques (Crossref metadata) Jimmy Lin, Xueguang Ma	primary paper abstract_only	ID: https://arxiv.org/bibtex/2106.14807 sha256:e1043019a7aa…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T14:18:00Z

Updated: 2026-10-19T14:18:00Z

Source File: data/curated/uniCOILUnifiedContextualizedInvertedList.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories