IndicRAGSuite: Large-Scale Datasets and a Benchmark for Indian Language RAG Systems

Anoop Kunchukuttan; Pasunuti Prasanjith; Prathmesh B More; Raj Dabre

REVIEW 6 major objections 6 minor 13 references

IndicRAGSuite: Large-Scale Datasets and a Benchmark for Indian Language RAG Systems

T0 review · 6 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a human-verified 13-language retrieval benchmark and a roughly 14-million-triplet synthetic training corpus close the resource gap for Indian-language retrieval-augmented generation.

desk verdict A genuinely useful resource paper for Indian-language RAG, but the 14M-triplet synthetic corpus needs quality evidence before it becomes the default training set. read the letter →

arxiv 2506.01615 v2 pith:DO2XBQFG submitted 2025-06-02 cs.CL

Pasunuti Prasanjith , Prathmesh B More , Anoop Kunchukuttan , Raj Dabre This is my paper

classification cs.CL

keywords retrieval-augmentedgenerationIndianlanguagesmultilingualinformationretrievalbenchmarkquestion-answer-reasoningtripletssynthetictrainingdatadenselow-resource

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

The paper sets out to fill two gaps that block retrieval-augmented generation for Indian languages: the absence of evaluation benchmarks and the scarcity of large training corpora. It introduces IndicMSMARCO, a benchmark of 1,000 machine-translated and then human-verified queries with passages in 13 Indian languages, so retrieval quality can be compared across languages and systems. For training, it contributes about 14 million question-answer-reasoning triplets generated from the Wikipedias of 19 Indian languages, plus translated versions of the standard MS MARCO training and development sets for 14 languages. If these resources hold up, developers of Indian-language RAG systems would no longer need to rely on English-centric datasets or translate queries on the fly, and researchers could measure progress with a standard ruler.

What carries the argument

The load-bearing mechanism is the question-answer-reasoning triplet: a cleaned Wikipedia paragraph is fed to an LLM, which is prompted to write up to three standalone questions, each with a short unambiguous answer and a reasoning span that quotes the paragraph. Keeping the paragraph as the unit, filtering out paragraphs that are too short or too long, and requiring reasoning grounded in the source are what make the synthetic corpus usable for training. The IndicMSMARCO benchmark is carried by translation-plus-verification: 1,000 queries and passages from the MS MARCO development set are machine-translated and then manually corrected for linguistic accuracy, semantic consistency, and named-entity fidelity.

What would settle it

Annotate a random sample of the Wikipedia-generated triplets for answer entailment and question answerability: if a meaningful share of the triplets fail, the dataset's quality claim is contradicted. Separately, train a dense retriever on the synthetic corpus alone and measure its MRR on IndicMSMARCO against the reported zero-shot baselines to test whether the corpus changes retrieval performance.

Watch

Extended reading notes

Core claim

The paper's central claim is that the missing infrastructure for Indian-language RAG can be built by pairing a small, carefully verified evaluation set with a very large, machine-generated training set. It reports that on IndicMSMARCO, existing dense retrieval models reach mean reciprocal rank scores of about 0.45 to 0.52 for most of the 13 languages, that the strongest model leads in 8 of 13 languages, and that Assamese and Odia score lowest. It also argues that translating queries and passages as whole paragraphs rather than sentence fragments preserves context and search intent, which makes the translated MS MARCO data more faithful than earlier Indian-language adaptations.

Load-bearing premise

The corpus's training value depends on the unverified premise that LLM-generated questions, answers, and reasoning spans drawn from Wikipedia paragraphs are accurate and faithful enough to learn from; the paper reports no human or automatic quality check on the 14-million-triplet portion.

Editorial extensions

If this is right

Retrieval models can now be compared on a common human-verified benchmark across 13 Indian languages, making evaluation standardized rather than ad hoc.
The 14-million-triplet Wikipedia corpus provides the scale needed for supervised dense-retriever training in Indian languages.
The full-paragraph translated MS MARCO sets add realistic search queries in 14 languages, avoiding the context fragmentation of sentence-level translation.
The reported baseline MRR scores give later work a concrete target to beat in each language, especially for low-resource Assamese and Odia.

Reading between the lines

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

The same paragraph-to-triplet generation recipe could be applied to any language with enough Wikipedia coverage, so the method is portable beyond the 19 languages listed in the paper.
The per-language score gaps invite a cross-lingual transfer experiment the paper does not run: train a retriever on high-resource Hindi, Telugu, or Malayalam data and test whether Assamese and Odia improve.
Because the paper stops at dataset construction and baseline evaluation, the 14-million-triplet corpus's actual value for training remains unconfirmed; the benchmark provides a ready test bed for that validation.

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

Worth a look if you work on Indian-language retrieval. The paper ships two things: a 13-language benchmark derived from 1000 MS MARCO dev queries, manually translated and post-edited, and a large training corpus of about 14M (question, answer, passage) triplets generated from Wikipedia in 19 languages, plus a paragraph-level translated MS MARCO. The benchmark fills a real gap: prior INDIC-MARCO had no human verification and only 11 languages. The scale is substantial, and the resources are public on HuggingFace. For an infrastructure paper, the contribution is real.

The soft spot is exactly where the reader's report puts it: there is no validation of the synthetic triplet dataset. Section 4.1.4 describes generation with Llama 3.3 70B and a filtering step based only on paragraph length. No pass rates, no human or automatic checks on answer correctness, no comparison against existing QA sets, no evidence that the reasoning spans actually support the answers. Because each positive passage is the paragraph the question was generated from, relevance is trivially true; the meaningful signal is whether the question-answer-reasoning triples are accurate and natural. The paper's own future work section concedes the absence of human-in-the-loop refinement. That is a load-bearing gap for a dataset whose main selling point is scale.

A second, more minor issue: the benchmark is claimed to be human-verified, but the paper gives no annotation protocol details, no inter-annotator agreement, and no examples of corrections. It may well be carefully done, but there is no evidence to assess it. Similarly, the retrieval experiments in Table 2 report MRR without retrieval setup details (corpus indexing, candidate count, passage chunking) or variance; these are baselines, not a rigorous benchmark study, so I would treat them as illustrative.

The MS MARCO translation itself is described as paragraph-level with IndicTrans3, and the authors correctly distance themselves from sentence-level INDIC-MARCO. That part looks sound from the description, though again no evaluation of translation quality.

Who is this for: anyone building or evaluating Indian-language retrieval/RAG systems. The benchmark is the clearer asset; the synthetic corpus could become useful but needs quality documentation. I'd send it to serious review with a request for validation experiments and annotation evidence, rather than desk reject.

Referee Report

6 major / 6 minor

Summary. The paper introduces IndicRAGSuite, a collection of resources intended to support retrieval-augmented generation for Indian languages. It comprises (1) IndicMSMARCO, a benchmark of 1,000 MS MARCO queries manually translated and post-edited into 13 Indian languages; (2) a synthetically generated Wikipedia-based dataset of roughly 14 million (question, answer, relevant passage) triplets across 19 languages, produced by prompting Llama 3.3 70B with paragraphs extracted from Wikipedia dumps; and (3) machine-translated versions of the MS MARCO train and dev sets into 14 Indian languages using IndicTrans3-beta. The paper reports dense retriever MRR scores on IndicMSMARCO for five off-the-shelf multilingual encoders and argues that these resources fill a gap in training and evaluation data for Indian-language retrieval.

Significance. If the resources are of adequate quality, IndicRAGSuite could be a useful contribution to multilingual IR for Indian languages. The scale alone is notable: 14 million synthetic triplets and a 1.3-million-query translated benchmark exceed existing Indian-language IR resources. The benchmark is derived from an external resource (MS MARCO) and evaluated with off-the-shelf models, so there is no fitted-parameter circularity. The synthetic Wikipedia data is generated by the same LLM family used for some translation steps, but this does not directly contaminate the evaluation. The datasets are publicly released on Hugging Face, which supports reproducibility and downstream use. However, the central claims of quality and usability currently rest on unverified generation and unquantified human verification, which limits confidence in the benchmark and dataset as reliable scientific artifacts.

major comments (6)

[§4.1.4–§4.1.5] The 14M-triplet Wikipedia dataset has no quality validation. Section 4.1.4 describes generation of questions, answers, and reasoning with Llama 3.3 70B, and Appendix A gives the prompt, but no results are reported from any automatic or human quality check: no answer-support verification, no pass rate for generated triplets, no assessment of question naturalness, no diversity metric, and no inter-annotator agreement. Because each 'relevant passage' is the paragraph used to generate the triplet, positive relevance is trivially true; the meaningful quality signals are whether answers are correct and grounded, and none are measured. The paper's own Section 4.3 lists 'human-in-the-loop refinement' as future work, acknowledging this absence. This gap is load-bearing for the training-data contribution, and it must be addressed before the dataset can be recommended for use.
[§3.2] The claim that IndicMSMARCO is 'human-verified' is not supported by any quantitative evidence. Section 3.2 describes three verification steps but gives no number of annotators per language, no inter-annotator agreement, no post-edit rate, and no examples of corrected errors. Without these, the reader cannot distinguish careful post-editing from a light review. At minimum, the authors should report agreement statistics or a sampled audit with error rates by language.
[§3.4, Table 2] The MRR results are not reproducible from the information given. The paper does not state the retrieval setup: the corpus used (translated passages? original English? Wikipedia?), the number of passages retrieved, the indexing method, the query preprocessing, or the pooling strategy. There are also no error bars, standard deviations, or statistical significance tests across languages or model runs. Since the benchmark is a central contribution, the evaluation protocol must be specified in enough detail for others to replicate the numbers.
[Abstract, §3.4, §5] The paper claims support for 'response generation' evaluation, but no generation experiments are reported. There are no RAG pipeline results, no generation metrics (e.g., ROUGE, BLEU, factual consistency), and no evaluation of generated answers. The benchmark may include passages useful for generation, but the paper never demonstrates that it can evaluate response generation. Either add generation experiments or remove the claim from the abstract and conclusion.
[Table 3, Table 4] The dataset statistics contain inconsistencies that need correction. Table 3 includes English (6.38M paragraphs before filtering) and labels the total as 19 languages, but the text consistently says '19 Indian languages'; English is not an Indian language, so the count and total are misleading. Table 4 includes Sanskrit among 14 languages, while the abstract and Section 3 say the benchmark covers 13 Indian languages; the relationship between the 13-language benchmark and the 14-language translated MS MARCO set needs clarification. Additionally, Marathi (200,000) and Urdu (199,999) before filtering are suspiciously round numbers that suggest placeholder or incomplete dumps; these should be verified and reported precisely.
[§4.2] The translated MS MARCO dataset is described as 'high-quality' based on the choice of IndicTrans3-beta and its 'superior BLEU scores', but no quality evaluation is presented for the actual translations. No BLEU/COMET scores, no human evaluation, and no comparison with the earlier IndicIRSuite/Indic-MARCO translations are reported. Since this dataset is a second major training resource, a quality assessment (even on a sample) is necessary to support the claim.

minor comments (6)

[§1, Key Contributions] Typo: '13 Indina languages' should be '13 Indian languages'.
[Figure 3] The figure caption references 'Figure ??' in the text; the reference should be resolved to the actual figure number.
[Figure 2] The Hindi example figure is hard to read in the supplied text; a higher-resolution rendering or transliteration would help readers verify the example.
[References] Several references contain formatting artifacts, such as 'V oyage' and 'Y ang', which should be corrected to 'Voyage' and 'Yang'.
[Table 3] The table would benefit from a column giving the fraction of paragraphs retained after filtering, since the retention rate varies widely (e.g., Kashmiri from 29,487 to 1,138) and the reasons for such drastic filtering are not discussed.
[§4.1.5] The filtering step excludes paragraphs that are 'too short' or 'too long', but no thresholds are given; the thresholds should be specified for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark is a translated subset of external MS MARCO, the evaluations use off-the-shelf retrievers, and no fitted parameter is presented as a prediction.

full rationale

The paper's two main artifacts are datasets, not derived results. IndicMSMARCO is constructed by sampling 1,000 queries and passages from the external MS MARCO dev set, translating with Llama 3.3 70B, and manually post-editing; the benchmark therefore inherits its relevance judgments from an external source rather than from the paper's own outputs. The evaluation in Table 2 applies pretrained dense retrievers directly with no fine-tuning or fitted parameters, so no 'prediction' reduces to a fitted input. The Wikipedia-generated triplets are produced by prompting Llama 3.3 70B over Wikipedia paragraphs; although the relevant passage is by construction the paragraph used for generation, the paper does not train or evaluate on this dataset, so this construction does not feed back into any claimed result. The only self-citation, IndicTrans3-beta (AI4Bharat), is used as a translation tool for the MS MARCO training set and is not invoked as a mathematical theorem or uniqueness argument; its quality is asserted rather than independently benchmarked, which is a rigor concern, not circularity. Section 4.3 explicitly lists human-in-the-loop refinement as future work, acknowledging the synthetic dataset has not been human-validated; that is a validity limitation, not a circular derivation. No equations or fitted parameters are present whose outputs equal their inputs by construction.

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

The central claims rest on the quality of machine translation and LLM-generated content, which are assumed rather than demonstrated. No free parameters or invented entities are introduced.

assumptions (4)

domain assumption MS MARCO is a reliable source of realistic queries and relevant passages for retrieval evaluation.
The benchmark is built from a subset of the MS MARCO dev set; if the source quality is low, the translation and evaluation inherit that (Section 3).
domain assumption Llama 3.3 70B generates fluent, accurate translations and question-answer triplets in 13 to 19 Indian languages.
Used for both benchmark translation (Section 3.1) and synthetic QA generation (Section 4.1.4) without reported human validation of outputs.
domain assumption Wikipedia paragraphs provide adequate grounding for diverse, coherent questions.
The pipeline accepts Wikipedia paragraph quality and topic coverage without analysis (Section 4.1.2).
domain assumption Paragraph-level translation preserves search intent better than sentence-level translation.
Assumes IndicTrans3-beta's paragraph-level output maintains semantic fidelity; no comparative evaluation is provided (Section 4.2).

how reviews work

0 comments

Cite this review

Pith. "Pith review of IndicRAGSuite: Large-Scale Datasets and a Benchmark for Indian Language RAG Systems." pith.science (2026). https://pith.science/paper/DO2XBQFG

@misc{pith2026250601615,
  author       = {Pith},
  title        = {Pith review of: IndicRAGSuite: Large-Scale Datasets and a Benchmark for Indian Language RAG Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DO2XBQFG}},
  note         = {Machine review of arXiv:2506.01615}
}

read the original abstract

Retrieval-Augmented Generation (RAG) systems enable language models to access relevant information and generate accurate, well-grounded, and contextually informed responses. However, for Indian languages, the development of high-quality RAG systems is hindered by the lack of two critical resources: (1) evaluation benchmarks for retrieval and generation tasks, and (2) large-scale training datasets for multilingual retrieval. Most existing benchmarks and datasets are centered around English or high-resource languages, making it difficult to extend RAG capabilities to the diverse linguistic landscape of India. To address the lack of evaluation benchmarks, we create IndicMSMarco, a multilingual benchmark for evaluating retrieval quality and response generation in 13 Indian languages, created via manual translation of 1000 diverse queries from MS MARCO-dev set. To address the need for training data, we build a large-scale dataset of (question, answer, relevant passage) tuples derived from the Wikipedias of 19 Indian languages using state-of-the-art LLMs. Additionally, we include translated versions of the original MS MARCO dataset to further enrich the training data and ensure alignment with real-world information-seeking tasks. Resources are available here: https://huggingface.co/collections/ai4bharat/indicragsuite-683e7273cb2337208c8c0fcb

Figures

Figures reproduced from arXiv: 2506.01615 by the authors.

**Figure 2.** Benchmark Example in Hindi 13 major Indian languages. Llama 3.3 70B ensures the retention of query intent while adapting languagespecific structures. • Passage Translation: The corresponding passages are translated using contextaware generation, ensuring coherence and fidelity to the original English passage. The model is prompted to preserve named entities, numerical data, and domainspecific terminology to … view at source ↗

**Figure 3.** The data processing pipeline—from raw Wikipedia dumps to paragraph extraction and an LLM—generating [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: System and task prompt used for generating highquality, languagespecific questionanswer pairs from [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 10 canonical work pages

[1]

the paragraph

Standalone & Additional ContextIndependent: The questions should be understandable without additional context and must not contain any references to “the paragraph” or “the article” outside of the content provided

work page
[2]

Unambiguous Answer: Each question should have a single, clear, and factual answer

work page
[3]

in relation to

Grounded in Context & Conceptual Format: Each question must be conceptually rooted in the provided article’s content and follow this format: Start with a clear question word (e.g., What, How, Where, When). Integrate key information from the article smoothly, using logical connectors (e.g., “in relation to”, “compared to”, “as a result of”, “which also...

work page
[4]

On the degree of semi-stable reduction

Scifact: A dataset for scientific claim verifica tion. In Proceedings of EMNLP. Shayne Longpre, Yi Lu, and Joachim Daiber. 2021. Mkqa: A multilingual knowledge question answering benchmark. arXiv preprint arXiv:2107.13613. Arvind Neelakantan, Tao Xu, Raul Puri, and 1 others

work page Pith review arXiv 2021
[7]

Bi-fidelity Reduced Polynomial Chaos Expansion for Uncertainty Quantification

Xtremer: Towards more challenging and multilingual multimodal learning. In Proceedings of EMNLP. Sebastian Ruder, Noah Constant, Jan Botha, and 1 oth ers. 2021. Xtremer: Towards more challenging and nuanced multilingual evaluation. arXiv preprint arXiv:2104.07462. V oyage AI. 2023. V oyagelite01instruct: Efficient multilingual embeddings. Technical ...

work page Pith review arXiv 2021
[8]

A high resolution study of near-IR diffuse interstellar bands, search for small scale structure, time variability and stellar features

E5: Towards text embeddings that transfer better across languages and tasks. arXiv preprint arXiv:2212.03563. Lee Xiong, Chenyan Xiong, Y e Li, and 1 others. 2021. Pretrained transformers for text ranking: Bert and beyond. In Proceedings of NAACLHLT. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami AlRfou, Aditya Siddhant, Aditya Barua, and Co...

work page Pith review arXiv 2021
[10]

tydi: A multilingual benchmark for dense retrieval

Mr. tydi: A multilingual benchmark for dense retrieval. In Proceedings of ACLIJCNLP. 9 Appendix A: Prompt Template for QuestionAnswerReasoning Generation from Wikipedia Articles System Prompt: You are a precise and helpful QuestionAnswer Generator that creates factual questions with verifiable answers from provided content in <target_language>. Task ...

work page
[2016]

arXiv preprint arXiv:1611.09268

Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268. Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceed ings of ACL. Yingqi Qu, Y uchen Ding, Jing Liu, and 1 others. 2021. Rocketqa: An optimized training approach to dense passage retrieval for opendomain qu...

arXiv 2019

Show all 13 references

[2018]

In Proceedings of EMNLP

Hotpotqa: A dataset for diverse, explainable multihop question answering. In Proceedings of EMNLP. Xinyu Zhang, Nandan Thakur, Barlas Oguz, Sachin Gupta, and Wentau Yih. 2023. Miracl: A multilin gual retrieval benchmark. In Proceedings of NeurIPS. Xinyu Zhang, Nandan Thakur...

2023
[2020]

In Proceedings of ACL

Crosslingual question answering as a starting point for zeroshot semantic parsing. In Proceedings of ACL. Akari Asai, Kyungjae Lee, Xing Li, and Eunsol Choi
[2021]

In Proceedings of ACLIJCNLP

Multilingual passage retrieval for opendomain question answering. In Proceedings of ACLIJCNLP. Luiz Bonifacio, Israel Campiotti, Rodrigo Nogueira, and Roberto Lotufo. 2021. mmarco: A multilingual version of ms marco passage ranking dataset. In Proceedings of EACL. Tanmoy Cha...

2021 arXiv
[2022]

OpenAI Technical Report

Text and code embeddings by contrastive pre training. OpenAI Technical Report. Anastasios Nentidis, Georgios Katsimpras, Anasta sia Krithara, Salvador Lima López, Eulália Farré Maduell, Luis Gasco, Martin Krallinger, and Geor gios Paliouras. 2023. Overview of BioASQ 2023: ...

2023
[2023]

In Proceedings of TREC

Neuclir: A benchmark for neural chinese language information retrieval. In Proceedings of TREC. Patrick Lewis, Barlas Oğuz, Ruty Rinott, Sebastian Riedel, and Holger Schwenk. 2020. Mlqa: Eval uating crosslingual extractive question answering . Preprint, arXiv:1910.07475. Ky...

2020 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

the paragraph

Standalone & Additional ContextIndependent: The questions should be understandable without additional context and must not contain any references to “the paragraph” or “the article” outside of the content provided

work page

[2] [2]

Unambiguous Answer: Each question should have a single, clear, and factual answer

work page

[3] [3]

in relation to

Grounded in Context & Conceptual Format: Each question must be conceptually rooted in the provided article’s content and follow this format: Start with a clear question word (e.g., What, How, Where, When). Integrate key information from the article smoothly, using logical connectors (e.g., “in relation to”, “compared to”, “as a result of”, “which also...

work page

[4] [4]

On the degree of semi-stable reduction

Scifact: A dataset for scientific claim verifica tion. In Proceedings of EMNLP. Shayne Longpre, Yi Lu, and Joachim Daiber. 2021. Mkqa: A multilingual knowledge question answering benchmark. arXiv preprint arXiv:2107.13613. Arvind Neelakantan, Tao Xu, Raul Puri, and 1 others

work page Pith review arXiv 2021

[5] [7]

Bi-fidelity Reduced Polynomial Chaos Expansion for Uncertainty Quantification

Xtremer: Towards more challenging and multilingual multimodal learning. In Proceedings of EMNLP. Sebastian Ruder, Noah Constant, Jan Botha, and 1 oth ers. 2021. Xtremer: Towards more challenging and nuanced multilingual evaluation. arXiv preprint arXiv:2104.07462. V oyage AI. 2023. V oyagelite01instruct: Efficient multilingual embeddings. Technical ...

work page Pith review arXiv 2021

[6] [8]

A high resolution study of near-IR diffuse interstellar bands, search for small scale structure, time variability and stellar features

E5: Towards text embeddings that transfer better across languages and tasks. arXiv preprint arXiv:2212.03563. Lee Xiong, Chenyan Xiong, Y e Li, and 1 others. 2021. Pretrained transformers for text ranking: Bert and beyond. In Proceedings of NAACLHLT. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami AlRfou, Aditya Siddhant, Aditya Barua, and Co...

work page Pith review arXiv 2021

[7] [10]

tydi: A multilingual benchmark for dense retrieval

Mr. tydi: A multilingual benchmark for dense retrieval. In Proceedings of ACLIJCNLP. 9 Appendix A: Prompt Template for QuestionAnswerReasoning Generation from Wikipedia Articles System Prompt: You are a precise and helpful QuestionAnswer Generator that creates factual questions with verifiable answers from provided content in <target_language>. Task ...

work page

[8] [2016]

arXiv preprint arXiv:1611.09268

Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268. Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceed ings of ACL. Yingqi Qu, Y uchen Ding, Jing Liu, and 1 others. 2021. Rocketqa: An optimized training approach to dense passage retrieval for opendomain qu...

arXiv 2019