Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Knowing a paper's concepts improves retrieval both in training data and at search time.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 13:00 UTC pith:QLYDKPVK

load-bearing objection Solid IR extension with a genuinely new inference-time snippet-matching component, but the paper understates its debt to the authors' own WSDM 2025 paper and leans on an unvalidated concept extractor. the 3 major comments →

arxiv 2601.00567 v2 pith:QLYDKPVK submitted 2026-01-02 cs.IR cs.AI

Improving Scientific Document Retrieval with Academic Concept Index

Jeyun Lee , Junhyoung Lee , Wonbin Kweon , Bowen Jin , Yu Zhang , Susik Yoon , Dongha Lee , Hwanjo Yu
show 2 more authors
Jiawei Han Seongku Kang
This is my paper
classification cs.IR cs.AI
keywords scientific document retrievalacademic concept indexsynthetic query generationconcept coveragecontext augmentationdense retrievalacademic taxonomyconcept extractor
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that LLM-based scientific retrieval fails to reach its potential because both of its main strategies—synthetic query generation and auxiliary context generation—ignore the conceptual structure inside a paper. To fix that, the authors build an academic concept index: per-document topics drawn from an academic taxonomy plus distinctive phrases, each weighted by importance. On that index, CCQGen generates a query set that progressively covers concepts the earlier queries missed, reducing redundancy; CCExpand produces concept-focused snippets and matches a query to the snippet whose conceptual focus best aligns with it. The experiments show consistent gains across metrics with both a general and a science-specific retriever, and the gains survive when the generated query budget is small. If correct, the index is a reusable, largely training-free component that can be stacked on top of existing retrieval enhancement methods.

Core claim

The central claim is that the limited improvement from LLM-based query generation and context augmentation in scientific retrieval is not a language-model problem but a conceptual-grounding problem. The solution is an academic concept index, built by selecting topics from an academic taxonomy and distinctive phrases from the corpus, then enriching both with importance estimates from a small concept extractor. CCQGen uses the index to detect which concepts previous queries left uncovered and conditions the LLM on those concepts, so successive queries complement rather than paraphrase one another. CCExpand uses the same index and the concept-aware queries to write short concept-focused snippet

What carries the argument

The carrying mechanism is the academic concept index plus the concept extractor that produces it. The extractor is a small multi-task classifier trained to predict LLM-selected core topics and phrases from a document, and its output distributions serve as importance weights. Those weights appear in three places: an under-coverage distribution drives which concepts the next generated query is conditioned on; a concept-aware similarity score is folded into round-trip consistency filtering; and a snippet-selection rule picks the concept-focused snippet most relevant to a test query. The taxonomy and phrase mining supply the candidate space; the extractor makes that space measurable and reusable

Load-bearing premise

The concept extractor's predicted phrase and topic probabilities are treated as a faithful measure of what a document is about and what a query already covers, but the paper never validates those probabilities against an external concept-labeling standard.

What would settle it

Replace the under-coverage distribution in CCQGen with uniform random sampling over the same enriched phrase candidates, keeping prompts, query budget, and filtering identical. If retrieval performance does not drop meaningfully, then the concept index's importance estimates are not what drives the reported gains, and the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Synthetic queries become more useful per query: consistent gains appear even when only a few queries are generated per document, easing compute and API-cost constraints.
  • Concept-aware training queries also improve plain document expansion, because the expansion inherits broader concept coverage before any snippet matching is applied.
  • The training-free side, CCExpand, can be deployed without retriever fine-tuning and with negligible online latency, since snippet generation and indexing happen offline and only top candidates are rescored.
  • The concept index composes with existing query-side and topic-based enhancement methods rather than replacing them, so its gains are partially additive.
  • Concept-aware filtering keeps queries that are lexically distant from the document but conceptually aligned, recovering valid training pairs that text-only round-trip filtering would discard.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The coverage mechanism is not retrieval-specific: the same under-coverage sampling could be used to build diverse evaluation sets, data augmentation, or summaries for scientific documents, wherever redundancy is the enemy.
  • Whether the taxonomy is essential is left untested; a corpus-derived phrase-only index would be the natural ablation to decide if the benefit comes from structure or simply from adding related terminology.
  • The central undeclared assumption—extractor calibration—could be probed by replacing the importance-weighted sampling in CCQGen with uniform random sampling over the same phrase candidates; if retrieval does not drop, the index's estimated importances are not the active ingredient.
  • Because the snippets are document-side and offline, the index could serve as a general infrastructure layer for scientific search, composable with future query-side expansions or rerankers.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces an academic concept index—a structured representation of each scientific document's key topics and phrases, enriched with importance weights and related terminology via a trained concept extractor. On top of this index, the authors propose CCQGen, which adaptively conditions LLM-based synthetic query generation on under-covered concepts, and CCExpand, which generates concept-focused snippets and uses the most aligned snippet as an auxiliary context during training-free retrieval. Experiments on CSFCube and DORIS-MAE with Contriever-MS and SPECTER-v2 report consistent improvements over several query generation and context augmentation baselines.

Significance. If validated, the proposed concept index would provide a reusable, corpus-level component for improving both fine-tuning data and inference-time matching in scientific document retrieval. The paper's strengths are the breadth of the evaluation—two datasets, two backbone retrievers, and multiple baseline families—and the inclusion of significance tests for the main CCQGen results (Table 1). The query-level analyses in Tables 4 and 5, the hyperparameter study in Figure 9, and the offline-snippet efficiency discussion are also useful. However, the central claim that concept coverage is the driving mechanism is not yet established because the concept extractor—the source of all concept signals—is unvalidated and appears in a self-referential loop. The per-dataset tuning of the CCExpand balance parameter without a validation split further weakens the evidential value of the reported gains.

major comments (3)
  1. [§3.2, Eq. (1), Eq. (4)] The concept extractor is the sole source of concept signals for both proposed methods, yet its accuracy is never checked against an external standard. It is trained on LLM-selected labels from Section 3.1; Eq. (1) uses its probabilities to define under-covered concepts; Eq. (4) uses its concept similarity as a filtering signal. This creates a self-referential loop: the same model decides which concepts are core, estimates query coverage, and scores query–document consistency. If the extractor is miscalibrated on short query text or the LLM labels contain systematic bias, CCQGen's sampling and filtering may simply reward agreement with the extractor rather than genuine conceptual coverage, and CCExpand inherits this dependency. Please (i) evaluate the extractor's precision/recall against human-annotated concept/keyphrase labels on a held-out set, and (ii) include an ablation using random
  2. [§6.1.5, §6.3.4] The CCExpand balance coefficient α is set to 0.6 for CSFCube and 0.4 for DORIS-MAE, but no separate validation split is described. Figure 9 reports a hyperparameter analysis only on DORIS-MAE with SPECTER-v2. If α was selected on the test sets, the reported improvements may be optimistic. Please state the selection procedure explicitly, use a validation split for both datasets, or show that fixed α across datasets preserves the main conclusions. In addition, Tables 6 and 7 report no standard deviations or significance tests; given that several baselines exhibit large swings across datasets, uncertainty quantification is needed to assess whether the CCExpand gains are stable.
  3. [Contributions vs. Ref. [16]] Reference [16] is titled 'Improving scientific document retrieval with concept coverage-based query set generation' and appears to describe the same CCQGen mechanism. The current paper lists CCQGen as a contribution without stating the relationship. Please clarify what is new relative to [16]: whether CCQGen is an already-published component, and if so, present the contribution as the unified academic concept index plus CCExpand, with CCQGen integrated as a building block. This is necessary for a fair assessment of novelty.
minor comments (5)
  1. [§6.1.1] The DORIS-MAE description says '165,144 test queries' without explaining the annotation process or the distribution of relevance labels. A sentence clarifying whether these are all used for evaluation would help, given that CSFCube has only 50 test queries.
  2. [§4.1, Eq. (1)] The use of ε=10^{-3} in max(ȳ_d^p − ȳ_Q^p, ε) assigns a small positive under-coverage probability to every core phrase even when the phrase is already covered. This is presumably for numerical stability, but it may also prevent exact zeroing of covered concepts. Please clarify the intended behavior and whether this choice affects the multinomial sampling in Eq. (2).
  3. [§4.2, Eq. (4)] The z-score normalization for s_text and s_concept is mentioned in the text, but the reference distribution used for the z-score (e.g., over what set of query–document pairs) is not specified. Please make this precise, as the filtering threshold and behavior depend on it.
  4. [§3.1.2] The distinctiveness formula exp(BM25(p,d))/(1+Σ_{d'∈D_d} exp(BM25(p,d'))) is not a probability and its numerical scale depends on BM25 scoring. The choice of top-20% distinctiveness and |D_d|=100 is also not justified. A brief rationale or sensitivity check would be helpful.
  5. [Figure 8] The efficiency plot would be more informative with error bars or at least the standard deviation of latency measurements, since LLM inference latency can be highly variable.

Circularity Check

0 steps flagged

No significant circularity: the concept index is a fitted component, but the retrieval claims are validated on external relevance labels.

full rationale

CCQGen and CCExpand are empirical methods rather than derivations from first principles. Eq. (1) does use the concept extractor's outputs for both document-level importance and query-level coverage, and Eq. (4) reuses the same extractor for consistency filtering; this creates a self-referential internal proxy for 'concept coverage' and would need external validation of the extractor. However, the headline results are NDCG, MAP, and Recall on CSFCube and DORIS-MAE, whose relevance labels are independent of the concept extractor's fitted values. Queries are generated and filtered before fine-tuning, and the final evaluation is on held-out test queries from these datasets. The self-citations [15], [16], and [17] provide related background and mechanisms from the same group, but they are not the sole justification for the experimental conclusions; comparisons are made against independent baselines (Promptgator, Pair-wise generation, GRF, HyDE) on external benchmarks. No step in the claimed derivation chain reduces by construction to its own inputs, so no significant circularity is established.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim leans on LLM-generated labels, the chosen taxonomy, and the concept extractor's probabilities. None of these are independently verified; the final NDCG labels are the only external anchors.

free parameters (3)
  • CCExpand balance alpha = 0.6 (CSFCube), 0.4 (DORIS-MAE)
    Balances document-level vs snippet-level similarity in Eq. 7; set per dataset with no separate validation split reported.
  • Number of queries/snippets M = 5
    Number of CCQGen queries and CCExpand snippets; described as empirically determined considering training resources, and default performance depends on it.
  • Concept index size hyperparameters = k_t=10, k_p=15, k_t'=15, k_p'=20; top-20% distinctiveness; |D_d|=100
    Hand-chosen limits for LLM core-topic/phrase selection and enrichment; the quality of the concept index depends on these choices.
axioms (4)
  • domain assumption The Microsoft Academic field-of-study taxonomy (431,416 nodes) is a meaningful topic hierarchy for scientific documents.
    Section 3.1.1 uses it to build the candidate topic set and top-down traversal; no sensitivity check to taxonomy choice is provided.
  • domain assumption LLM judgments for topic/phrase selection and snippet generation are sufficiently accurate and free of hallucination.
    Section 3.1 restricts candidates to reduce hallucination but does not measure residual errors; Section 5.1 relies on LLM snippets for matching.
  • domain assumption The concept extractor's predicted probabilities measure true concept importance and query coverage.
    Section 3.2 trains the extractor; Eq. 1 and Eq. 4 use its outputs as the coverage signal, with no external evaluation of extractor accuracy.
  • domain assumption CSFCube and DORIS-MAE relevance judgments are valid ground truth.
    Section 6.1.1 treats the datasets as given; label noise and annotation procedure differences are not discussed.

pith-pipeline@v1.3.0-alltime-deepseek · 23864 in / 12631 out tokens · 123555 ms · 2026-08-03T13:00:48.982224+00:00 · methodology

0 comments
read the original abstract

Adapting general-domain retrievers to scientific domains is challenging due to the scarcity of large-scale domain-specific relevance annotations and the substantial mismatch in vocabulary and information needs. Recent approaches address these issues through two independent directions that leverage large language models (LLMs): (1) generating synthetic queries for fine-tuning, and (2) generating auxiliary contexts to support relevance matching. However, both directions overlook the diverse academic concepts embedded within scientific documents, often producing redundant or conceptually narrow queries and contexts. To address this limitation, we introduce an academic concept index, which extracts key concepts from papers and organizes them guided by an academic taxonomy. This structured index serves as a foundation for improving both directions. First, we enhance the synthetic query generation with concept coverage-based generation (CCQGen), which adaptively conditions LLMs on uncovered concepts to generate complementary queries with broader concept coverage. Second, we strengthen the context augmentation with concept-focused auxiliary contexts (CCExpand), which leverages a set of document snippets that serve as concise responses to the concept-aware CCQGen queries. Extensive experiments show that incorporating the academic concept index into both query generation and context augmentation leads to higher-quality queries, better conceptual alignment, and improved retrieval performance.

Figures

Figures reproduced from arXiv: 2601.00567 by Bowen Jin, Dongha Lee, HwanJo Yu, Jeyun Lee, Jiawei Han, Junhyoung Lee, SeongKu Kang, Susik Yoon, Wonbin Kweon, Yu Zhang.

Figure 1
Figure 1. Figure 1: Overview of our framework. (a) We construct an academic concept index by extracting key concepts [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: The overview of Concept Coverage-based Query set Generation (CCQGen). Best viewed in color. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An illustration of concept-focused snippet generation in CCExpand. Best viewed in color. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: An illustration of concept-focused relevance matching process. Best viewed in color. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results with varying amounts of training data. x% denotes setups using a random x% of queries. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Improvements by concept coverage-based filtering. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Trade-off between retrieval effectiveness and efficiency. Blue and orange markers denote Contriever [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Hyperparameter analysis: (a) 𝛼 (with fixed 𝑀 = 5), and (b) 𝑀 (with fixed 𝛼 = 0.6). 6.3.4 Hyperparameter study. To guide the hyperparameter selection of CCExpand, we conduct a detailed analysis on DORIS-MAE using SPECTER-v2 as the backbone model. In [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 7 linked inside Pith

  1. [1]

    Chris Alberti, Daniel Andor, Emily Pitler, Jacob Devlin, and Michael Collins. 2019. Synthetic QA Corpora Generation with Roundtrip Consistency. InACL. 6168–6173

    2019

  2. [2]

    Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: Pretrained Language Model for Scientific Text. InEMNLP. 3615–3620

    2019

  3. [3]

    Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Nogueira. 2022. Inpars: Unsupervised dataset generation for information retrieval. InSIGIR. 2387–2392

    2022

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. InNeurIPS, Vol. 33. 1877–1901

    2020

  5. [5]

    Aditi Chaudhary, Karthik Raman, and Michael Bendersky. 2024. It’s All Relative!–A Synthetic Query Generation Approach for Improving Zero-Shot Relevance Prediction. InFindings of NAACL

    2024

  6. [6]

    Aditi Chaudhary, Karthik Raman, Krishna Srinivasan, Kazuma Hashimoto, Mike Bendersky, and Marc Najork. 2023. Exploring the viability of synthetic query generation for relevance prediction. InThe SIGIR 2023 Workshop on eCom- merce

    2023

  7. [7]

    Qin Chen, Qinmin Hu, Jimmy Xiangji Huang, and Liang He. 2018. Modeling queries with contextual snippets for information retrieval.ACM Transactions on Intelligent Systems and Technology9, 4 (2018), 1–26

    2018

  8. [8]

    Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, and Daniel S. Weld. 2020. SPECTER: Document-level Representation Learning using Citation-informed Transformers. InACL

    2020

  9. [9]

    Zhuyun Dai, Vincent Y Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B Hall, and Ming-Wei Chang. 2023. Promptagator: Few-shot dense retrieval from 8 examples. InICLR

    2023

  10. [10]

    Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and Stéphane Clinchant. 2022. From distillation to hard negative sampling: Making sparse neural ir models more effective. InSIGIR. 2353–2359

    2022

  11. [11]

    Luyu Gao and Jamie Callan. 2021. Condenser: a Pre-training Architecture for Dense Retrieval. InEMNLP

    2021

  12. [12]

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. Precise zero-shot dense retrieval without relevance labels. InACL. 1762–1777

    2023

  13. [13]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning.arXiv preprint arXiv:2112.09118 (2021)

    Pith/arXiv arXiv 2021

  14. [14]

    Vitor Jeronymo, Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, Roberto Lotufo, Jakub Zavrel, and Rodrigo Nogueira

  15. [15]

    SeongKu Kang, Shivam Agarwal, Bowen Jin, Dongha Lee, Hwanjo Yu, and Jiawei Han. 2024. Improving Retrieval in Theme-specific Applications using a Corpus Topical Taxonomy. InWWW. 1497–1508

    2024

  16. [16]

    SeongKu Kang, Bowen Jin, Wonbin Kweon, Yu Zhang, Dongha Lee, Jiawei Han, and Hwanjo Yu. 2025. Improving scientific document retrieval with concept coverage-based query set generation. InWSDM. 895–904

    2025

  17. [17]

    SeongKu Kang, Yunyi Zhang, Pengcheng Jiang, Dongha Lee, Jiawei Han, and Hwanjo Yu. 2024. Taxonomy-guided Semantic Indexing for Academic Paper Search. InEMNLP. 7169–7184

    2024

  18. [18]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open-Domain Question Answering. InEMNLP. 6769–6781

    2020

  19. [19]

    Dongha Lee, Jiaming Shen, SeongKu Kang, Susik Yoon, Jiawei Han, and Hwanjo Yu. 2022. Taxocom: Topic taxonomy completion with hierarchical discovery of novel topic clusters. InWWW. 2819–2829

    2022

  20. [20]

    Haitao Li, Qingyao Ai, Jia Chen, Qian Dong, Yueyue Wu, Yiqun Liu, Chong Chen, and Qi Tian. 2023. SAILER: Structure-aware Pre-trained Language Model for Legal Case Retrieval. InSIGIR

    2023

  21. [21]

    Yunzhe Li, Qian Chen, Weixiang Yan, Wen Wang, Qinglin Zhang, and Hari Sundaram. 2024. Advancing Precise Outline-Conditioned Text Generation with Task Duality and Explicit Outline Control. InEACL. 2362–2377

    2024

  22. [22]

    Davis Liang, Peng Xu, Siamak Shakeri, Cicero Nogueira dos Santos, Ramesh Nallapati, Zhiheng Huang, and Bing Xiang. 2020. Embedding-based zero-shot retrieval through query generation.arXiv preprint arXiv:2009.10270(2020)

    Pith/arXiv arXiv 2020

  23. [23]

    Xiao Liu, Da Yin, Jingnan Zheng, Xingjian Zhang, Peng Zhang, Hongxia Yang, Yuxiao Dong, and Jie Tang. 2022. OAG-BERT: Towards a Unified Backbone Language Model for Academic Knowledge Services. InKDD. 3418–3428

    2022

  24. [24]

    Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Daniel S Weld. 2020. S2ORC: The Semantic Scholar Open Research Corpus. InACL. 4969–4983

    2020

  25. [25]

    Ji Ma, Ivan Korotkov, Yinfei Yang, Keith Hall, and Ryan McDonald. 2021. Zero-shot Neural Passage Retrieval via Domain-targeted Synthetic Question Generation. InEACL. 1075–1088

    2021

  26. [26]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. 2018. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. InKDD. 1930–1939. , Vol. 1, No. 1, Article . Publication date: January 2026. 24 Jeyun Lee and Junhyoung Lee et al

    2018

  27. [27]

    Iain Mackie, Shubham Chatterjee, and Jeffrey Dalton. 2023. Generative Relevance Feedback with Large Language Models. InSIGIR. 2026–2031

    2023

  28. [28]

    Watheq Mansour, Shengyao Zhuang, Guido Zuccon, and Joel Mackenzie. 2024. Revisiting document expansion and filtering for effective first-stage retrieval. InSIGIR. 186–196

    2024

  29. [29]

    Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Generation- Augmented Retrieval for Open-Domain Question Answering. InACL. 4089–4100

    2021

  30. [30]

    Sheshera Mysore, Arman Cohan, and Tom Hope. 2022. Multi-Vector Models with Textual Guidance for Fine-Grained Scientific Document Similarity. InNAACL. 4453–4470

    2022

  31. [31]

    Sheshera Mysore, Tim O’Gorman, Andrew McCallum, and Hamed Zamani. 2021. CSFCube-A Test Collection of Computer Science Research Articles for Faceted Query by Example.NeurIPS 2021 Track on Datasets and Benchmarks (2021)

    2021

  32. [32]

    Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. 2019. From doc2query to docTTTTTquery.Online preprint6, 2 (2019)

    2019

  33. [33]

    Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document expansion by query prediction.arXiv preprint arXiv:1904.08375(2019)

    Pith/arXiv arXiv 2019

  34. [34]

    Malte Ostendorff, Nils Rethmeier, Isabelle Augenstein, Bela Gipp, and Georg Rehm. 2022. Neighborhood Contrastive Learning for Scientific Document Representations with Citation Embeddings. InEMNLP

    2022

  35. [35]

    Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang

  36. [36]

    Anastasiia Razdaibiedina and Aleksandr Brechalov. 2023. MIReAD: Simple Method for Learning High-quality Repre- sentations from Scientific Documents. InACL. 530–539

    2023

  37. [37]

    Jon Saad-Falcon, Omar Khattab, Keshav Santhanam, Radu Florian, Martin Franz, Salim Roukos, Avirup Sil, Md Sultan, and Christopher Potts. 2023. UDAPDR: Unsupervised Domain Adaptation via LLM Prompting and Distillation of Rerankers. InEMNLP. 11265–11279

    2023

  38. [38]

    Devendra Sachan, Mike Lewis, Mandar Joshi, Armen Aghajanyan, Wen-tau Yih, Joelle Pineau, and Luke Zettlemoyer

  39. [39]

    Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207(2021)

    Pith/arXiv arXiv 2021

  40. [40]

    Sonal Sannigrahi, Thiago Fraga-Silva, Youssef Oualil, and Christophe Van Gysel. 2024. Synthetic query generation using large language models for virtual assistants. InSIGIR. 2837–2841

    2024

  41. [41]

    Jingbo Shang, Jialu Liu, Meng Jiang, Xiang Ren, Clare R Voss, and Jiawei Han. 2018. Automated phrase mining from massive text corpora.IEEE Transactions on Knowledge and Data Engineering30, 10 (2018), 1825–1837

    2018

  42. [42]

    Zhihong Shen, Hao Ma, and Kuansan Wang. 2018. A Web-scale system for scientific knowledge exploration. InACL. 87–92

    2018

  43. [43]

    Amanpreet Singh, Mike D’Arcy, Arman Cohan, Doug Downey, and Sergey Feldman. 2023. SciRepEval: A Multi-Format Benchmark for Scientific Document Representations. InEMNLP. 5548–5566

    2023

  44. [44]

    Fangbo Tao, Honglei Zhuang, Chi Wang Yu, Qi Wang, Taylor Cassidy, Lance M Kaplan, Clare R Voss, and Jiawei Han

  45. [45]

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models. InNeurIPS Datasets and Benchmarks Track

    2021

  46. [46]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971(2023)

    Pith/arXiv arXiv 2023

  47. [47]

    Jianyou Wang, Kaicheng Wang, Xiaoyue Wang, Prudhviraj Naidu, Leon Bergen, and Ramamohan Paturi. 2023. Scientific Document Retrieval using Multi-level Aspect-based Queries. InNeurIPS Datasets and Benchmarks Track

    2023

  48. [48]

    Kexin Wang, Nandan Thakur, Nils Reimers, and Iryna Gurevych. 2022. GPL: Generative Pseudo Labeling for Unsuper- vised Domain Adaptation of Dense Retrieval. InNAACL. 2345–2360

    2022

  49. [49]

    Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query Expansion with Large Language Models. InEMNLP. 9414–9423

    2023

  50. [50]

    Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners.arXiv preprint arXiv:2109.01652(2021)

    Pith/arXiv arXiv 2021

  51. [51]

    Lingling Xu, Haoran Xie, Zongxi Li, Fu Lee Wang, Weiming Wang, and Qing Li. 2023. Contrastive Learning Models for Sentence Representations: A Survey.ACM Transactions on Intelligent Systems and Technology14, 4 (2023), 1–34

    2023

  52. [52]

    HongChien Yu, Chenyan Xiong, and Jamie Callan. 2021. Improving query representations for dense retrieval with pseudo relevance feedback. InCIKM. 3592–3596. , Vol. 1, No. 1, Article . Publication date: January 2026. Improving Scientific Document Retrieval with Academic Concept Index 25

    2021

  53. [53]

    Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. InSIGIR. 1503–1512

    2021

  54. [54]

    Hang Zhang, Yeyun Gong, Yelong Shen, Jiancheng Lv, Nan Duan, and Weizhu Chen. 2022. Adversarial Retriever-Ranker model for Dense Retrieval. InICLR

    2022

  55. [55]

    Yu Zhang, Hao Cheng, Zhihong Shen, Xiaodong Liu, Ye-Yi Wang, and Jianfeng Gao. 2023. Pre-training Multi-task Contrastive Learning Models for Scientific Literature Understanding. InFindings of EMNLP. 12259–12275

    2023

  56. [56]

    Wangchunshu Zhou, Yuchen Eleanor Jiang, Ethan Wilcox, Ryan Cotterell, and Mrinmaya Sachan. 2023. Controlled text generation with natural language instructions. InICML. 42602–42613. , Vol. 1, No. 1, Article . Publication date: January 2026

    2023

  57. [2016]

    Bull.39, 3 (2016), 74–84

    Multi-Dimensional, Phrase-Based Summarization in Text Cubes.IEEE Data Eng. Bull.39, 3 (2016), 74–84

    2016

  58. [2021]

    InNAACL-HLT

    RocketQA: An Optimized Training Approach to Dense Passage Retrieval for Open-Domain Question Answering. InNAACL-HLT. 5835–5847

  59. [2022]

    Improving Passage Retrieval with Zero-Shot Question Generation. InEMNLP. 3781–3797

  60. [2023]

    Inpars-v2: Large language models as efficient dataset generators for information retrieval.arXiv preprint arXiv:2301.01820(2023)

    Pith/arXiv arXiv 2023