Home / Curated / Algorithms / InPars In-Context Data Augmentation for Information Retrieval

InPars In-Context Data Augmentation for Information Retrieval

Algorithms AIMech:InParsInContextDataAugmentationForInformationRetrieval DRAFT

Aliases: InPars, Inquisitive Parrots for Search, InPars Few-Shot Synthetic Query Generation for IR

A few-shot synthetic data augmentation method for information retrieval (Inquisitive Parrots for Search) that prompts a pretrained language model with a fixed prefix of N (e.g., 3) document-question demonstrations—either standard pairs (Vanilla) or pairs contrasting a generic 'bad question' with a descriptive 'good question' (Guided by Bad Questions, GBQ)—to generate a synthetic question for each sampled corpus document, filters the generated pairs by selecting the top K with highest length-normalized token log-probability p_q = (1/|q|) sum_{i=1}^{|q|} log p(q_i | t, d, q_{<i}), samples negative documents from BM25 candidates, and fine-tunes a neural reranker.

Identity Notes & Concept Scope Distinct from Promptagator (AIMech:PromptagatorFewShotDenseRetrievalDataGeneration), which filters synthetic pairs via round-trip top-1 retrieval consistency using an initial dual encoder and trains a first-stage dense dual encoder, whereas InPars filters synthetic pairs directly by their length-normalized autoregressive generation log-probability p_q under the generator LM and trains a monoT5 sequence-to-sequence reranker over BM25 candidates.

Claims & Grounded Evidence

4 Claims

Mechanism #inpars-length-normalized-logprob-filtering-mechanism

Given a few-shot prompt prefix t and sampled document d, InPars autoregressively generates a candidate question q and retains only the top K (e.g., 10,000 out of 100,000) synthetic (q, d) pairs ranked by average token log-probability p_q = (1/|q|) sum_{i=1}^{|q|} log p(q_i | t, d, q_{<i}).

supports Source: InPars: Data Augmentation for Information Retrieval using Large Language Models · Locator: Section 3 & Section 6.3, pp. 2-3 & 8, Equation 1

Mechanism #inpars-guided-by-bad-questions-prompting-mechanism

To prevent the language model from generating overly generic questions that match many passages, the Guided by Bad Questions (GBQ) template pairs each in-context passage with both a shallow MS MARCO 'bad question' and a context-specific 'good question', instructing the model to generate a good question for the target document.

supports Source: InPars: Data Augmentation for Information Retrieval using Large Language Models · Locator: Section 4.2 & Section 6.1, pp. 4 & 7, Figure 2

Application #inpars-neural-reranker-adaptation-application

InPars is applied to unsupervised and domain-adapted passage and document reranking on MS MARCO, TREC-DL 2020, Robust04, Natural Questions, and TREC-COVID, enabling monoT5 rerankers trained without human-labeled in-domain queries to outperform BM25, Contriever, and direct GPT-3 API reranking.

supports Source: InPars: Data Augmentation for Information Retrieval using Large Language Models · Locator: Section 5, p. 6, Table 1
supports Source: InPars: Data Augmentation for Information Retrieval using Large Language Models (Crossref metadata) · Locator: DBLP bibliographic record Luiz Henrique Bonifacio et al. SIGIR 2022

Limitation #inpars-first-stage-bm25-recall-dependency-limitation

Because InPars trains a cross-encoder reranker applied to the top-1000 documents retrieved by bag-of-words BM25, its overall recall is upper-bounded by lexical first-stage retrieval and it can slightly degrade models already heavily supervised on in-domain MS MARCO queries.

supports Source: InPars: Data Augmentation for Information Retrieval using Large Language Models · Locator: Section 4.3 & Section 5, pp. 5-7

⚠️ Curated Limitations & Negative Results

Generating 100,000 candidate questions per target collection requires substantial LLM inference calls before filtering.
Randomly sampling negatives from the top-1000 BM25 results can occasionally introduce false-negative passages.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	related_match	InPars uses a few-shot prompted neural language model to generate synthetic query-document training data for neural transformer rerankers.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r741-primary-pdf`	InPars: Data Augmentation for Information Retrieval using Large Language Models Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, Rodrigo Nogueira	primary paper full_text	ID: https://dblp.org/search/publ/api?q=InPars+Data+Augmentation+for+Information+Retrieval+Bonifacio&h=5&format=xml sha256:b2e1e7be17b7…
`r741-arxiv-abs`	InPars: Data Augmentation for Information Retrieval using Large Language Models (arXiv abstract) Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, Rodrigo Nogueira	primary paper abstract_only	ID: https://arxiv.org/abs/2202.05144 sha256:2106e22d85f2…
`r741-metadata-cr`	InPars: Data Augmentation for Information Retrieval using Large Language Models (Crossref metadata) Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, Rodrigo Nogueira	primary paper abstract_only	ID: https://dblp.org/search/publ/api?q=InPars+Data+Augmentation+for+Information+Retrieval+Bonifacio&h=5&format=xml sha256:f8bb31831600…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T13:44:00Z

Updated: 2026-10-19T13:44:00Z

Source File: data/curated/InParsInContextDataAugmentationForInformationRetrieval.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