Pith. sign in

REVIEW 4 major objections 4 minor 49 references

KG-CQR: Leveraging Structured Relation Representations in Knowledge Graphs for Contextual Query Retrieval

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Knowledge-graph contexts lift RAG retrieval accuracy on hard queries without retraining

desk verdict A sensible incremental KG-query-enrichment pipeline with plausible but statistically unproven gains; the tables contradict themselves, so treat the headline numbers as provisional. read the letter →

arxiv 2508.20417 v3 pith:M5ZLVEC4 submitted 2025-08-28 cs.CL cs.DB

classification cs.CLcs.DB
keywords contextualqueryretrievalretrieval-augmentedgenerationknowledgegraphexpansionsubgraphextractionmulti-hopquestionansweringtextualtripletrepresentation
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

This paper proposes KG-CQR, a training-free pipeline that enriches a user's query with a concise, knowledge-graph-derived context before embedding it for retrieval in retrieval-augmented generation (RAG) systems. The central claim is that converting a corpus into a knowledge graph, extracting a query-relevant subgraph, completing it with inferred paths, and having an LLM summarize that subgraph into a 'contextual query' markedly improves retrieval metrics on multi-hop and domain-specific benchmarks. If correct, the approach offers a model-agnostic way to bridge the query-document embedding gap without fine-tuning or extra supervision.

What carries the argument

The central mechanism is the Textual Triplet Representation (TTR): each knowledge-graph triple (head, relation, tail) is expanded by an LLM into a natural-language sentence that describes the relation. Subgraph extraction scores triples by cosine similarity between the query embedding and the TTR embeddings (Equation 6), which moves matching from entity/keyword level to sentence level. The subgraph completion step then adds paths between extracted entities using beam-search-guided BFS, scoring paths by mean TTR-query similarity, and the contextual generation step summarizes the completed subgraph into a single query context. The final retrieval vector is a weighted fusion of the original que

What would settle it

Run KG-CQR on RAGBench or MultiHop-RAG using a KG constructed by a small or intentionally corrupted LLM (or by injecting NER errors into the triplets), and compare retrieval mAP: if the mAP gain over the plain BGE baseline vanishes or reverses, the method's benefit is confirmed to hinge on KG construction quality rather than on query enrichment per se.

Watch

Extended reading notes

Core claim

KG-CQR treats the retrieval problem as a context-generation problem: instead of matching the raw query against documents, it first builds a corpus-centric knowledge graph, retrieves a small set of triples whose textual triplet representations (TTRs) are most similar to the query, filters erroneous triples with an LLM, completes the subgraph by beam-search path addition, and then has an LLM generate a short contextual summary of that subgraph. This summary is embedded in the document space and linearly fused with the original query embedding. The paper reports that on RAGBench and MultiHop-RAG, KG-CQR yields 4–6% relative mAP gains and 2–3% relative Recall@25 gains over strong sparse and dens

Load-bearing premise

The whole pipeline depends on the knowledge graph built by LLaMA-3.3-70B being accurate, and on the LLM-generated textual triplet summaries faithfully representing the triples—if entity extraction or relation extraction is wrong, the contextual query is wrong before retrieval even starts.

Editorial extensions

If this is right

  • RAG systems can adopt KG-CQR as a drop-in query-enrichment layer for existing sparse or dense retrievers, achieving retrieval gains without retraining the retriever.
  • Multi-hop and domain-specific retrieval tasks that suffer from query-document embedding misalignment benefit most, since the KG-derived context supplies intermediate entities and relations the raw query omits.
  • The approach reduces the number of reasoning iterations in multi-step RAG pipelines, because the contextual query already contains the bridging evidence needed to answer a multi-hop question.
  • Because the pipeline is model-agnostic, the same KG-CQR module can be paired with small LLMs (e.g., 3B parameters) and still retain most of the retrieval benefit, lowering the compute barrier for practical deployment.
  • Combining KG-CQR with corpus-centric KG methods (e.g., HippoRAG2) is complementary, improving retrieval metrics further than either approach alone.

Reading between the lines

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

  • A testable extension would be replacing the LLM-generated TTRs with a lighter extractive summarizer (e.g., selecting a sentence from the source document that mentions the same head and tail), to test whether the TTR bonus comes from LLM paraphrase quality or from simply having sentence-level query alignment.
  • Since the paper's own error analysis shows failures on fine-grained temporal reasoning and comparative queries, a natural next step is adding a timestamp or numeric-attribute channel to the TTR representation, which could be done without changing the subgraph-completion machinery.
  • If the method's gains come mainly from the fusion weight (α = 0.7), then a learned or query-adaptive α could squeeze further gains; this is not explored in the paper.
  • The paper reports retrieval metrics only; a direct comparison of end-to-end answer faithfulness under KG-CQR versus HyDE would clarify whether retrieval gains translate into fewer hallucinations, which is the stated motivation for the work.
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.

Referee Report

4 major / 4 minor

Summary. KG-CQR is a training-free pipeline that enriches queries for retrieval in RAG systems. It constructs a corpus-centric knowledge graph with LLaMA-3.3-70B, augments each triplet with a Textual Triplet Representation (TTR), extracts query-relevant triplets via cosine similarity and LLM filtering, completes the subgraph using beam-search BFS path addition, and generates a contextual query summary that is fused with the original query by a weighted sum (Eq. 9). Retrieval is then performed by encoding the fused representation with the document encoder. The authors report evaluations on RAGBench and MultiHop-RAG with BM25, DPR, and BGE backbones, plus multi-step RAG experiments with IRCoT on RAGBench, HotpotQA, and MuSiQue. They claim consistent gains of 4–6% mAP and 2–3% Recall@25 over strong baselines, model-agnostic scalability, and improved multi-hop reasoning performance.

Significance. If the reported gains are reliable, KG-CQR is a practical, training-free query-enrichment method with a plausible mechanism: structured triplets and textual triplet representations can supply corpus-grounded context that a raw query lacks. The paper's strengths include detailed algorithmic descriptions (Algorithms 1 and 2), prompt templates in Appendix 6.5, an ablation study, a latency comparison, and qualitative error analysis. The framework's model-agnostic design and the absence of training make it easy to adopt. However, the central quantitative claim is not yet statistically grounded: results are single-run point estimates, the fusion weight α is tuned on the test benchmarks, and one multi-hop experiment has internally inconsistent baseline numbers. These issues are fixable but currently limit confidence in the headline conclusions.

major comments (4)
  1. [Section 4.2, Table 1] All retrieval numbers are single-run point estimates without error bars, confidence intervals, or significance tests. The pipeline is stochastic at multiple stages: KG construction, TTR generation, LLM filtering (Eq. 7), and contextual generation (Eq. 8) all involve LLM sampling. On MultiHop-RAG the BGE improvements are mAP 0.227→0.240 and Recall@25 0.520→0.525; these deltas are small enough that run-to-run variance from the LLM components could plausibly reverse them. The claim 'KG-CQR significantly improves retrieval performance' therefore lacks statistical support. Repeated runs with multiple seeds and significance testing, or at least a variance report, are needed.
  2. [Appendix 6.2.2, Table 7, Eq. (9)] The fusion weight α is selected by sweeping α ∈ {0.3, 0.5, 0.7} on the same RAGBench and MultiHop-RAG test sets used for the final evaluation, and α=0.7 is then reported as the best configuration. This makes the headline numbers test-set-tuned rather than out-of-sample. A proper validation split (or nested selection) should be used to select α, and the final numbers should be reported on a held-out test set. The same concern applies to other hyperparameters (k, K, beam width W, max path length L), for which no sensitivity analysis is provided.
  3. [Section 4.3.3, Table 4 vs. Appendix 6.2.1, Table 6] There is an internal contradiction in the multi-step RAG results. For MuSiQue with LLaMA-3.3-70B, Table 4 reports BM25 retrieval F1 = 0.374, while Table 6 lists the baseline as 0.479. If the second value is actually BGE (the appendix says 'using BGE as the retrieval baseline'), the row labels in the HotpotQA and MuSiQue panels of Table 6 are wrong, because they say 'BM25' and 'KG-CQR + BM25.' As printed, the two tables report different values for an identically labeled configuration, which undermines the reliability of the multi-hop RAG claims. Please correct the tables and clarify which retriever each number corresponds to.
  4. [Abstract and Section 4.2] The abstract's claim of 'approximately 2–3% improvement in Recall@25' is only supported by the RAGBench numbers (Recall@25 0.657→0.675, about +2.7% relative). On MultiHop-RAG, the corresponding improvement is 0.520→0.525, about +1% relative, which is within plausible noise given the stochastic pipeline. The stated range should be qualified per dataset, or the MultiHop-RAG claim should be withdrawn until variance is measured. The mAP claim of 4–6% is also based on single-run estimates and would need the same statistical backing.
minor comments (4)
  1. [Appendix 6.2.1, Table 6] The table headers and row labels are inconsistent: the RAGBench panel correctly uses 'BGE' and 'KG-CQR + BGE,' but the HotpotQA and MuSiQue panels say 'BM25' and 'KG-CQR + BM25,' even though the appendix text states that BGE is the retrieval baseline. This formatting issue makes the appendix hard to interpret.
  2. [Equation (6)] The notation 'argmax_{T_i^KG ∈ TKG, k} sim(v_q, v_i^r)' is nonstandard; it appears to mean 'select the top-k triples by similarity.' Please state this explicitly, e.g., 'top-k selection.'
  3. [Algorithm 1] Step 3 in the text says paths containing nodes not present in the initial subgraph are filtered, but the pseudocode (lines 18–22) does not implement this filtering. Please align the description and the code.
  4. [Figure 4] The latency comparison reports relative values without units or a description of the experimental environment. Adding absolute latencies (e.g., seconds per query) and the hardware used would make the scalability claim more concrete.

Circularity Check

1 steps flagged · score 5.0 of 10

Headline retrieval gains are partially fitted: the fusion weight α is selected on the same test sets used to report the improvements.

  1. fitted input called prediction [Section 4.2 (Table 1) and Appendix 6.2.2 (Table 7)]
    "The reported results use α= 0.7 (Equation 9), which was found to yield the best performance (the selection of this value is further discussed in Appendix 6.2.2). ... Table 7 shows the comprehensive evaluation on the value of α to fuse the input query and context embeddings (Equation 9). As results, setting α= 0.7 consistently yields the best overall performance."

    The fusion weight α in Equation 9 is a hyperparameter selected by sweeping over the RAGBench and MultiHop-RAG test sets (Appendix 6.2.2, Table 7). The same test-set numbers are then presented as the headline results in Table 1 (e.g., KG-CQR + BGE mAP 0.542, Recall@25 0.675). Thus the reported 'improvement' is not an out-of-sample prediction; it is the best test-set performance over the α grid. The central claim 'KG-CQR outperforms strong baselines by 4–6% mAP and 2–3% Recall@25' is therefore partly a fitted result rather than a parameter-free validation, reducing the strength of the empirical derivation.

full rationale

The paper does not exhibit definitional circularity: KG-CQR is a concrete pipeline (subgraph extraction via TTR similarity, completion via beam-search BFS, contextual generation, and weighted fusion), and the retrieval numbers are measured from actual runs, not derived from the method's definition. There is no load-bearing self-citation: the cited works by the same research group (Bui et al. 2021; Pham et al. 2025a,b) support general claims about domain sensitivity and ambiguity, not the core mechanism or the uniqueness of KG-CQR. Likewise, no ansatz is smuggled in via self-citation; GraphRAG is cited only for the standard corpus-centric KG construction procedure. The main circularity concern is the test-set selection of α. Equation 9 defines the final query representation as a weighted sum of the original query embedding and the KG-CQR context embedding. The paper explicitly states that α=0.7 'was found to yield the best performance' and the appendix reports the full α sweep on the same two benchmarks. Consequently, the headline mAP and Recall@25 gains are the result of optimizing α on the evaluation data, so the reported superiority is in-sample by construction. This fits the 'fitted input called prediction' pattern: a parameter is fit to the test distribution, and the resulting test numbers are then presented as evidence of the method's effectiveness. Beyond this, the MultiHop-RAG improvements are small (e.g., Recall@25 0.520→0.525, +0.005) and the paper reports no variance or multi-seed statistics, so the gains may not be statistically robust. However, that is a correctness/statistical concern rather than a circularity concern. The limitations section honestly acknowledges KG-construction errors and limited evaluation scope, which further supports that the method is not being defined into success. Overall, the central derivation is not circular, but the headline empirical claim is partially fitted, yielding a score of 5.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The system relies on free hyperparameters (alpha, K, beam width, path length) tuned for the reported results, and on domain assumptions about KG quality, TTR fidelity, and the doc-space encoding of generated text. No new physical or conceptual entities are introduced.

free parameters (5)
  • fusion weight α = 0.7
    Weighted sum in Eq. 9; chosen as best over {0.3, 0.5, 0.7} on the test sets (Appendix 6.2.2).
  • top-k triplets for subgraph extraction = not reported
    k in Eq. 6 controls the number of initial triplets; value not justified.
  • K for subgraph completion = 20
    Top-K unique triplets added in Algorithm 1 Step 4.
  • beam width W = 3
    Beam Search width in Algorithm 2.
  • max path length L = not specified
    Limits path length in Algorithm 2.
assumptions (3)
  • domain assumption The LLM-built corpus-centric KG accurately represents the facts in the corpus.
    Section 3.1.1 and Limitations: KG construction via LLaMA-3.3-70B is susceptible to NER/RE/EL errors that propagate.
  • domain assumption TTR embeddings capture query-relevant semantics of triplets.
    Equations 2 and 6 rely on cosine similarity between query embedding and TTR embeddings; no validation that TTRs are faithful.
  • domain assumption The generated context text lies in the document embedding space.
    Equation 5 follows HyDE (Gao et al., 2023); assumes the LLM-generated summary is encoded well by the document encoder.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KG-CQR: Leveraging Structured Relation Representations in Knowledge Graphs for Contextual Query Retrieval." pith.science (2026). https://pith.science/paper/M5ZLVEC4

@misc{pith2026250820417,
  author       = {Pith},
  title        = {Pith review of: KG-CQR: Leveraging Structured Relation Representations in Knowledge Graphs for Contextual Query Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M5ZLVEC4}},
  note         = {Machine review of arXiv:2508.20417}
}
read the original abstract

The integration of knowledge graphs (KGs) with large language models (LLMs) offers significant potential to improve the retrieval phase of retrieval-augmented generation (RAG) systems. In this study, we propose KG-CQR, a novel framework for Contextual Query Retrieval (CQR) that enhances the retrieval phase by enriching the contextual representation of complex input queries using a corpus-centric KG. Unlike existing methods that primarily address corpus-level context loss, KG-CQR focuses on query enrichment through structured relation representations, extracting and completing relevant KG subgraphs to generate semantically rich query contexts. Comprising subgraph extraction, completion, and contextual generation modules, KG-CQR operates as a model-agnostic pipeline, ensuring scalability across LLMs of varying sizes without additional training. Experimental results on RAGBench and MultiHop-RAG datasets demonstrate KG-CQR's superior performance, achieving a 4-6% improvement in mAP and a 2-3% improvement in Recall@25 over strong baseline models. Furthermore, evaluations on challenging RAG tasks such as multi-hop question answering show that, by incorporating KG-CQR, the performance consistently outperforms the existing baseline in terms of retrieval effectiveness

Figures

Figures reproduced from arXiv: 2508.20417 by the authors.

Figure 1
Figure 1. Overview of query expansion approaches for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Construction of structured relation represen [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An illustration of KG-CQR for the retrieval process, which includes three main components: Subgraph [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Retrieval latency 8B). Similar improvements are observed on Hot￾potQA, where KG-CQR yields a significant gain for the largest model (F1 = 0.700 vs. 0.663). The effect is most pronounced on MuSiQue, where KG￾CQR + BM25 achieves F1 = 0.489 with LLaMA￾3.3-70B compared to …
Figure 5
Figure 5. Figure 5: Prompt the task instruction for KG construction [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Prompt the task instruction for textual triplet representation. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Prompt the task instruction for filtering irrelevant triplets [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Prompt the task instruction for contextual representation [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 18 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Hasan Abu - Rasheed, Christian Weber, and Madjid Fathi. 2024. https://doi.org/10.1109/EDUCON60312.2024.10578654 Knowledge graphs as context sources for llm-based explanations of learning recommendations . In IEEE Global Engineering Education Conference, EDUCON 2024, Kos Island, Greece, May 8-11, 2024 , pages 1--5. IEEE

  4. [4]

    Hiteshwar Kumar Azad and Akshay Deepak. 2019. https://doi.org/10.1016/J.IPM.2019.05.009 Query expansion techniques for information retrieval: A survey . Inf. Process. Manag., 56(5):1698--1735

  5. [5]

    Do, Yan Xu, and Pascale Fung

    Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. https://doi.org/10.18653/V1/2023.IJCNLP-MAIN.45 A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity . In Proceedings of the 13th...

  6. [6]

    Sebastian Bruch, Siyu Gai, and Amir Ingber. 2024. https://doi.org/10.1145/3596512 An analysis of fusion functions for hybrid retrieval . ACM Trans. Inf. Syst. , 42(1):20:1--20:35

  7. [7]

    Manh - Ha Bui, Toan Tran, Anh Tran, and Dinh Q. Phung. 2021. https://proceedings.neurips.cc/paper/2021/hash/b0f2ad44d26e1a6f244201fe0fd864d1-Abstract.html Exploiting domain-specific features to enhance domain generalization . In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2...

  8. [8]

    Chi - Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, and Jie Fu. 2024. https://doi.org/10.48550/ARXIV.2404.00610 RQ-RAG: learning to refine queries for retrieval augmented generation . CoRR, abs/2404.00610

Show all 49 references
  1. [9]

    Xinran Chen, Xuanang Chen, Ben He, Tengfei Wen, and Le Sun. 2024. https://doi.org/10.18653/V1/2024.FINDINGS-ACL.708 Analyze, generate and refine: Query expansion with llms for zero-shot open-domain QA . In Findings of the Association for Computational Linguistics, ACL 2024, Ba...

  2. [10]

    Wentao Ding, Jinmao Li, Liangchuan Luo, and Yuzhong Qu. 2024. https://doi.org/10.1145/3589334.3645563 Enhancing complex question answering over knowledge graphs through evidence pattern retrieval . In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-1...

  3. [11]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. 2024. https://doi.org/10.48550/ARXIV.2404.16130 From local to global: A graph RAG approach to query-focused summarization . CoRR, abs/2404.16130

  4. [12]

    Robert Friel, Masha Belyi, and Atindriyo Sanyal. 2024. https://doi.org/10.48550/ARXIV.2407.11005 Ragbench: Explainable benchmark for retrieval-augmented generation systems . CoRR, abs/2407.11005

  5. [13]

    Jinlan Fu, See - Kiong Ng, Zhengbao Jiang, and Pengfei Liu. 2024. https://doi.org/10.18653/V1/2024.NAACL-LONG.365 Gptscore: Evaluate as you desire . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lang...

  6. [14]

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. https://doi.org/10.18653/V1/2023.ACL-LONG.99 Precise zero-shot dense retrieval without relevance labels . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers...

  7. [15]

    Bernal Jimenez Gutierrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. http://papers.nips.cc/paper\_files/paper/2024/hash/6ddc001d07ca4f319af96a3024f6dbd1-Abstract-Conference.html Hipporag: Neurobiologically inspired long-term memory for large language models . In A...

  8. [16]

    Bernal Jim \' e nez Guti \' e rrez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. 2025. https://doi.org/10.48550/ARXIV.2502.14802 From RAG to memory: Non-parametric continual learning for large language models . CoRR, abs/2502.14802

  9. [17]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick S. H. Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen - tau Yih. 2020. https://doi.org/10.18653/V1/2020.EMNLP-MAIN.550 Dense passage retrieval for open-domain question answering . In Proceedings of the 2020 Conference on...

  10. [18]

    Xiaoxi Li, Jiajie Jin, Yujia Zhou, Yuyao Zhang, Peitian Zhang, Yutao Zhu, and Zhicheng Dou. 2025. https://doi.org/10.1145/3722552 From matching to generation: A survey on generative information retrieval . ACM Trans. Inf. Syst. Just Accepted

  11. [19]

    Xingxuan Li, Ruochen Zhao, Yew Ken Chia, Bosheng Ding, Shafiq Joty, Soujanya Poria, and Lidong Bing. 2024. https://openreview.net/forum?id=cPgh4gWZlz Chain-of-knowledge: Grounding large language models via dynamic knowledge adapting over heterogeneous sources . In The Twelfth ...

  12. [20]

    Yanming Liu, Xinyue Peng, Jiannan Cao, Shi Bo, Yanxin Shen, Xuhong Zhang, Sheng Cheng, Xun Wang, Jianwei Yin, and Tianyu Du. 2025. https://openreview.net/pdf?id=cPozlf9OaF Bridging context gaps: Leveraging coreference resolution for long contextual understanding . In The Thirt...

  13. [21]

    Yaojie Lu, Qing Liu, Dai Dai, Xinyan Xiao, Hongyu Lin, Xianpei Han, Le Sun, and Hua Wu. 2022. https://doi.org/10.18653/V1/2022.ACL-LONG.395 Unified structure generation for universal information extraction . In Proceedings of the 60th Annual Meeting of the Association for Comp...

  14. [22]

    Linhao Luo, Yuan - Fang Li, Gholamreza Haffari, and Shirui Pan. 2024. https://openreview.net/forum?id=ZGNWW7xZ6Q Reasoning on graphs: Faithful and interpretable large language model reasoning . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vie...

  15. [23]

    Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.322 Query rewriting in retrieval-augmented large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 20...

  16. [24]

    Shengyu Mao, Yong Jiang, Boli Chen, Xiao Li, Peng Wang, Xinyu Wang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. 2024. https://aclanthology.org/2024.findings-emnlp.49 Rafe: Ranking feedback improves query rewriting for RAG . In Findings of the Association for Computa...

  17. [25]

    Morris and Alexander M

    John X. Morris and Alexander M. Rush. 2024. https://doi.org/10.48550/ARXIV.2410.02525 Contextual document embeddings . CoRR, abs/2410.02525

  18. [26]

    Thi Nguyen, Linhao Luo, Fatemeh Shiri, Dinh Phung, Yuan - Fang Li, Thuy - Trang Vu, and Gholamreza Haffari. 2024. https://doi.org/10.18653/V1/2024.FINDINGS-ACL.168 Direct evaluation of chain-of-thought in multi-hop reasoning with knowledge graphs . In Findings of the Associati...

  19. [27]

    Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu. 2024. https://doi.org/10.1109/TKDE.2024.3352100 Unifying large language models and knowledge graphs: A roadmap . IEEE Trans. Knowl. Data Eng. , 36(7):3580--3599

  20. [28]

    Hoang Pham, Thanh-Do Nguyen, and Khac-Hoai Nam Bui. 2025 a . https://aclanthology.org/2025.findings-acl.274/ C laim PKG : Enhancing claim verification via pseudo-subgraph generation with lightweight specialized LLM . In Findings of the Association for Computational Linguistics...

  21. [29]

    Hoang Pham, Thanh-Do Nguyen, and Khac-Hoai Nam Bui. 2025 b . https://aclanthology.org/2025.naacl-long.268/ Verify-in-the-graph: Entity disambiguation enhancement for complex claim verification with interactive graph representation . In Proceedings of the 2025 Conference of the...

  22. [30]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. https://jmlr.org/papers/v21/20-074.html Exploring the limits of transfer learning with a unified text-to-text transformer . J. Mach. Learn. Res.,...

  23. [31]

    Robertson and Hugo Zaragoza

    Stephen E. Robertson and Hugo Zaragoza. 2009. https://doi.org/10.1561/1500000019 The probabilistic relevance framework: BM25 and beyond . Found. Trends Inf. Retr., 3(4):333--389

  24. [32]

    Keshav Santhanam, Omar Khattab, Jon Saad - Falcon, Christopher Potts, and Matei Zaharia. 2022. https://doi.org/10.18653/V1/2022.NAACL-MAIN.272 Colbertv2: Effective and efficient retrieval via lightweight late interaction . In Proceedings of the 2022 Conference of the North Ame...

  25. [33]

    Bhaskarjit Sarmah, Dhagash Mehta, Benika Hall, Rohan Rao, Sunil Patel, and Stefano Pasquali. 2024. https://doi.org/10.1145/3677052.3698671 Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction . In Proceedings of...

  26. [34]

    Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D. Manning. 2024. https://openreview.net/forum?id=GN921JHCRw RAPTOR: recursive abstractive processing for tree-organized retrieval . In The Twelfth International Conference on Learning Repres...

  27. [35]

    Ni, Heung - Yeung Shum, and Jian Guo

    Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel M. Ni, Heung - Yeung Shum, and Jian Guo. 2024. https://openreview.net/forum?id=nnVO1PvbTv Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph . In The Tw...

  28. [36]

    Yixuan Tang and Yi Yang. 2024. https://doi.org/10.48550/ARXIV.2401.15391 Multihop-rag: Benchmarking retrieval-augmented generation for multi-hop queries . CoRR, abs/2401.15391

  29. [37]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. https://doi.org/10.1162/TACL\_A\_00475 Musique: Multihop questions via single-hop question composition . Trans. Assoc. Comput. Linguistics, 10:539--554

  30. [38]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. https://doi.org/10.18653/V1/2023.ACL-LONG.557 Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions . In Proceedings of the 61st Annual Meeting of th...

  31. [39]

    Liang Wang, Nan Yang, and Furu Wei. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.585 Query2doc: Query expansion with large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 20...

  32. [40]

    Yuqi Wang, Boran Jiang, Yi Luo, Dawei He, Peng Cheng, and Liangcai Gao. 2024 a . https://doi.org/10.48550/ARXIV.2404.10384 Reasoning on efficient knowledge paths: Knowledge graph guides large language model for domain question answering . CoRR, abs/2404.10384

  33. [41]

    Yuxia Wang, Minghan Wang, Muhammad Arslan Manzoor, Fei Liu, Georgi Georgiev, Rocktim Jyoti Das, and Preslav Nakov. 2024 b . https://aclanthology.org/2024.emnlp-main.1088 Factuality of large language models: A survey . In Proceedings of the 2024 Conference on Empirical Methods ...

  34. [42]

    Rossi, Haoliang Wang, and Julian McAuley

    Yu Xia, Junda Wu, Sungchul Kim, Tong Yu, Ryan A. Rossi, Haoliang Wang, and Julian McAuley. 2025. https://aclanthology.org/2025.naacl-long.216/ Knowledge-aware query expansion with large language models for textual and relational retrieval . In Proceedings of the 2025 Conferenc...

  35. [43]

    Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian - Yun Nie. 2024. https://doi.org/10.1145/3626772.3657878 C-pack: Packed resources for general chinese embeddings . In Proceedings of the 47th International ACM SIGIR Conference on Research and Devel...

  36. [44]

    Yao Xu, Shizhu He, Jiabei Chen, Zihao Wang, Yangqiu Song, Hanghang Tong, Guang Liu, Jun Zhao, and Kang Liu. 2024. https://aclanthology.org/2024.emnlp-main.1023 Generate-on-graph: Treat LLM as both agent and KG for incomplete knowledge graph question answering . In Proceedings ...

  37. [45]

    Cohen, Ruslan Salakhutdinov, and Christopher D

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. https://doi.org/10.18653/V1/D18-1259 Hotpotqa: A dataset for diverse, explainable multi-hop question answering . In Proceedings of the 2018 Conference...

  38. [46]

    Hongbin Ye, Ningyu Zhang, Hui Chen, and Huajun Chen. 2022. https://doi.org/10.18653/V1/2022.EMNLP-MAIN.1 Generative knowledge graph construction: A review . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United...

  39. [47]

    Le Zhang, Yihong Wu, Qian Yang, and Jian - Yun Nie. 2024. https://aclanthology.org/2024.findings-emnlp.103 Exploring the best practices of query expansion with large language models . In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA...

  40. [48]

    Shunyu Zhang, Yaobo Liang, Ming Gong, Daxin Jiang, and Nan Duan. 2022. https://doi.org/10.18653/V1/2022.ACL-LONG.414 Multi-view document representation learning for open-domain dense retrieval . In Proceedings of the 60th Annual Meeting of the Association for Computational Lin...

  41. [49]

    Yuqi Zhu, Xiaohan Wang, Jing Chen, Shuofei Qiao, Yixin Ou, Yunzhi Yao, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2024. https://doi.org/10.1007/S11280-024-01297-W Llms for knowledge graph construction and reasoning: recent capabilities and future opportunities . World Wide We...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.