Pith. sign in

REVIEW 5 major objections 7 minor 47 references

Aligning Knowledge Graphs and Language Models for Factual Accuracy

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that aligning knowledge-graph entity embeddings with an LLM's text embedding space, via a trainable projection layer, improves factual question answering and helps the model tell similar entities apart.

desk verdict Useful incremental result on infusing KGE embeddings into LLMs for single-entity KG-QA, but the hallucination-reduction claim is not supported by the experiments and the evaluation has post-hoc selection and compute-mismatch issues. read the letter →

arxiv 2507.13411 v1 pith:OZLMPDPZ submitted 2025-07-17 cs.CL cs.AI

Nur A Zarin Nishat , Andrea Coletta , Luigi Bellomarini , Kossi Amouzouvi , Jens Lehmann , Sahar Vahdati This is my paper
classification cs.CLcs.AI
keywords knowledgegraphembeddingLLMhallucinationmitigationentity-textalignmenttrainableprojectionlayerquestionansweringTransEparameter-efficientfine-tuningfactualgrounding
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

ALIGNed-LLM tries to establish that a large language model's factual accuracy can be improved by injecting structured knowledge-graph information directly into the model's latent text space, without retrieving triples into the prompt and without full retraining. The approach takes a frozen knowledge-graph embedding of the question's reference entity, projects that vector into the language model's embedding space with a lightweight trainable layer, and concatenates it with the question's token embeddings before generation. If the central claim is right, this alignment gives the model a richer entity-specific context, so it can distinguish similar entities (for example two companies with near-identical names) and ground answers in KG facts rather than stale or confabulated memory. The paper reports consistent improvements in exact match and related metrics over fine-tuned baselines across three open QA benchmarks and an anonymized central-bank company-ownership use case.

What carries the argument

The load-bearing mechanism is the projection layer $\phi$ that maps a frozen KGE entity embedding $X_e \in \mathbb{R}^{d_e}$ into the LLM's token-embedding space $\mathbb{R}^{d_q}$, followed by concatenation $H = [H_q : H_e]$ with the query's token embeddings. The paper implements $\phi$ as an identity map, a linear map, or an MLP with GELU activations depending on the KG and model; the choice affects results, with linear projection working best on the sparse company-ownership graph. The KGE side is TransE, whose translation property $X_h + r \approx X_t$ makes the entity vectors encode relational structure, and training proceeds in two stages: first align $\phi$ with the LLM frozen, then fine-tune $\phi$ and the LLM head together with a cross-entropy loss. This is what lets the model see entity identity and graph position while generating, rather than only seeing a textual name.

What would settle it

Replace the correct entity embedding in $H$ with a random KG entity embedding while keeping the question unchanged; if exact match does not drop substantially relative to using the correct entity, the reported gains are not caused by faithful entity grounding.

Watch

Extended reading notes

Core claim

The paper's central discovery is that pre-trained knowledge-graph embeddings and LLM text embeddings can be aligned closely enough with a single trainable projection layer that the LLM's generation becomes measurably more factual. Concretely, ALIGNed-LLM looks up the entity embedding $X_e$ for the reference entity, maps it through $\phi$ into the text embedding space to obtain $H_e = \phi(X_e)$, and forms the LM input as the concatenation $H = [H_q : H_e]$, maximizing $P(X_a | X_q, X_e)$ over generated tokens. Because $X_e$ comes from a KGE model like TransE trained on the KG, the vector carries relational structure (for instance $X_h + r \approx X_t$), giving the language model information about where the entity sits in the graph. The experiments show that this infusion consistently improves exact-match, ROUGE, BLEU, and F1 scores over a fine-tuned baseline on Wikidata, YAGO3-10, Mintaka, and the company-ownership dataset, with the largest gains for smaller models such as TinyLlama and Mistral.

Load-bearing premise

The pipeline assumes that each question comes with a known single entity whose KG embedding can be looked up; for arbitrary user questions this requires an external entity linker that the paper neither implements nor evaluates.

Editorial extensions

If this is right

  • Knowledge-grounded factuality can be improved with a lightweight two-stage fine-tuning, avoiding full retraining and avoiding the need to linearize KG triples into the prompt.
  • Knowledge updates become cheap: the paper states that updating the KG requires only updating the entity embeddings, not retraining the language model.
  • The approach transfers to high-accuracy enterprise settings such as banking ownership-control questions, where ALIGNed models improve exact-match answers over fine-tuning alone.
  • The best projection architecture depends on KG topology, so practitioners should choose between linear and complex projections based on the graph's structure.

Reading between the lines

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

  • The paper assumes a single known entity per query and never evaluates entity linking, so applying the pipeline to arbitrary user questions requires an external linker whose errors could erase the gains.
  • A direct comparison with retrieval-augmented prompting using the same triples is not run; such a test would show whether latent-space infusion adds anything beyond giving the model the same facts as text.
  • If the projection layer learns a general semantic alignment, it could be reused for entity disambiguation or hallucination detection outside QA; probing it on entities outside the fine-tuning set would test that.
  • The alignment idea extends naturally to temporal or multi-modal knowledge graphs, which the paper lists only as future 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

5 major / 7 minor

Summary. The paper proposes ALIGNed-LLM, a method for infusing knowledge graph (KG) information into an LLM's latent space by projecting frozen KG entity embeddings through a trainable projection layer and concatenating the projected entity representation with the query's token embeddings. The model is trained in two stages: projection-layer pretraining followed by joint end-to-end fine-tuning with the LLM head. Experiments are reported on Wikidata, YAGO3-10, Mintaka, and a proprietary anonymized company-ownership (CO) dataset, using Vicuna-13B, Mistral-7B, and TinyLlama-1.1B, with Exact Match, ROUGE, BLEU, and F1 metrics. The central claim is that this alignment improves factual grounding and reduces hallucination, based on consistent gains over fine-tuned baselines.

Significance. If the claims are substantiated, the method is a simple and computationally lightweight way to inject structured KG knowledge into decoder-only LLMs, and the central-bank use case gives it practical relevance. The paper provides a useful transparency package: full per-metric tables, statistical significance tests, query templates, and a small error analysis. However, the contribution is incremental relative to prior KG-infusion methods such as KnowLA and KG-Adapter, which are discussed only in related work and not compared empirically. The claimed hallucination reduction is not directly measured, and several experimental choices limit confidence in the headline gains. The method is best viewed as a contribution to single-entity, KG-derived question answering rather than a general solution to LLM factuality.

major comments (5)
  1. [§IV.A and Appendix Table VIII] The projection type (linear vs. complex) is selected per dataset after observing test-set results: Table VIII states that the complex projection is used for all models except ALIGNed TL-1.1B on CO, where a linear projection is used 'to achieve best result.' Because the same test set is used for selection and reporting, the reported gains are partially the result of post-hoc model selection. Please select the projection type on a validation split or pre-register the choice per model/dataset, and report performance for both projection types on the test set.
  2. [§III.D and §IV] The comparison is not compute-matched: ALIGNed-LLM receives an additional projection-pretraining stage (stage one) before end-to-end fine-tuning, whereas the fine-tuned baselines receive only the end-to-end stage. The observed improvements could therefore be due to the extra training step or the additional trainable parameters, rather than to the KG-derived entity embeddings. Please add an ablation that gives the baseline the same amount of additional training (e.g., extra fine-tuning with LoRA) or that trains the projection layer with randomly initialized or scrambled entity embeddings while keeping the pipeline identical.
  3. [Abstract, §IV, and §V] The abstract and conclusion claim that ALIGNed-LLM 'reduces hallucination,' but no experiment measures hallucination as ungrounded or unsupported generation. Section IV evaluates only answer-overlap metrics (EM, ROUGE, BLEU, F1) on KG-derived QA pairs, and the error analysis in Appendix F covers only 200 examples in the company-ownership setting. These metrics do not support a general hallucination-reduction claim. Please either add a hallucination-oriented evaluation (e.g., open-ended generation with fact-verification or human annotation) or explicitly narrow the claim to 'improves accuracy on single-entity KG QA where the reference entity is provided.'
  4. [§III and §IV.A] The method depends on a strong assumption stated in Section III: 'we assume that the query is constructed around a single entity e that is provided along with the instruction.' Every dataset is structured as (Xh, Xq, Xt), so the entity embedding is looked up from the gold head entity and no entity linking is evaluated. This means the reported gains measure the model's ability to use a provided entity embedding to retrieve the tail entity, not general factual reasoning or open-domain QA. Please state this limitation prominently in the main text and, ideally, evaluate at least one setting with an external entity linker or with missing/wrong entities to quantify the method's applicability.
  5. [§II and §IV] The related-work section describes closely related methods such as KnowLA, KG-Adapter, and ELPF, which also infuse KG representations into LLMs, but the experiments compare only against fine-tuned LLM baselines. The claim that ALIGNed-LLM is 'simple yet effective' would be substantially strengthened by an empirical comparison with at least one prior KG-infusion baseline under the same training and inference conditions. Without such a comparison, the novelty and relative advantage over existing approaches are not demonstrated.
minor comments (7)
  1. [§III.D] There is a typo in the subsection title: 'ALINGNed-LLM' should be 'ALIGNed-LLM'.
  2. [Appendix A] The name 'Tabale V' should be 'Table V', and 'anoynmized' should be 'anonymized'.
  3. [§IV, Evaluation Metrics] The definition of 'Reverse Weighted BLEU (RWB)' is unclear: the text describes weighting lower-order n-grams but does not specify the exact formula or how it differs from standard BLEU. Please provide a precise definition or a citation.
  4. [§IV, Baselines] The phrase 'the fine-tuned version of each LLM' implicitly uses LoRA/PEFT, but the method section says 'fine-tuning' without specifying that the baseline also uses LoRA. Please clarify the exact parameter-efficient fine-tuning setup used for the baselines so the comparison is unambiguous.
  5. [Appendix B] No code or data link is provided. Given the reproducibility value of the anonymized CO dataset and the exact query templates, please include an availability statement or a repository link.
  6. [Table II] The table reports 'Wikidata5M' as a dataset but also lists separate Wikidata QA and Mintaka datasets; the label is confusing because Wikidata5M is the KG, not a QA dataset. Please label the KG and the QA datasets separately.
  7. [Appendix F] The error-analysis tables are based on 200 examples per model; the percentages in Tables XIV and XV should be accompanied by confidence intervals or at least the total count of each error type, since small-sample percentages can be noisy.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation-level circularity: ALIGNed-LLM is a standard supervised alignment-and-fine-tuning pipeline whose stated equations do not reduce to their inputs.

full rationale

I walked the claimed derivation chain from Section III through the experiments. The method takes a frozen KGE entity embedding Xe, maps it through a trainable projection layer (Eqs. 2-5), concatenates the result with the query embedding (Eq. 6), and optimizes a cross-entropy generation loss (Eq. 8). Nothing in these equations defines the target answer in terms of the input projection or vice versa; the projection is learned from paired QA examples, so the setup is a conventional supervised learning pipeline. The related-work citations to Baldazzi et al. [4, 5] are self-citations from the Banca d'Italia group, but they are used only as domain motivation and are not load-bearing for the ALIGNed-LLM claim. I also checked the two places where circularity most often hides in this genre. First, the QA task is defined as predicting tail Xt from head Xh and question Xq, and the KGE is trained on 'the reference KG'; if the KGE train split were not disjoint from the QA test triples, then the head embedding could carry in-sample answer information and the reported gains would partly reflect embedding leakage. However, the paper reports no explicit statement that test triples were used to train TransE, and the baselines are fine-tuned on the same QA pairs, so this is a potential evaluation-scope or leakage concern rather than a derivation equivalence established by the paper's own equations. Second, the abstract's 'reducing hallucination' claim is not directly measured by EM, ROUGE, BLEU, or F1 on KG-derived QA, and the error analysis is limited to 200 examples; that is a validity and scope limitation, not a circular derivation. Under the strict definition of circularity used here, the derivation chain is self-contained: the projection parameters and LLM are fit to the same objective they are evaluated on, which is the normal supervised setting, so the circularity score is 0.

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

The paper rests on empirical design choices rather than a derivation. No new real-world entities are introduced. The main free parameters are the projection architecture (selected post hoc), the projection training length, and the embedding dimension/source. The axioms concern entity availability, the equivalence of KG-QA accuracy with hallucination reduction, and the sufficiency of TransE embeddings for disambiguation.

free parameters (4)
  • Projection layer type (linear vs complex) = linear for ALIGNed TL-1.1B on CO; complex for all other reported configurations
    The choice was made after observing test results; Table VIII caption states linear was used 'to achieve best result' for TL-1.1B on CO.
  • Projection training epochs = 50 for YAGO3-10 and CO, 100 for Wikidata and Mintaka
    Appendix B reports per-dataset choices without a validation-based selection procedure.
  • Entity embedding dimension = 512 for Wikidata5M, 1024 for YAGO3-10 and CO
    Taken from pre-trained models; the paper tests 512 vs 1024 on CO and reports small differences (Table XIII).
  • TransE implementation = GraphVite for Wikidata, RotatE framework for others
    Different KGE training pipelines are used per dataset; the effect of this choice on alignment is not controlled.
assumptions (4)
  • domain assumption Each QA question is associated with exactly one reference entity e whose KG embedding is available.
    Stated in Section III: 'we assume that the query is constructed around a single entity e that is provided along with the instruction.' The pipeline depends on this assumption.
  • domain assumption The QA pairs generated from KG triples are a valid measure of the model's factual accuracy and hallucination.
    The paper evaluates only on KG-derived questions (Wikidata, YAGO, Mintaka, CO), not on open-domain hallucination benchmarks; the abstract's hallucination claim relies on this equivalence.
  • domain assumption TransE embeddings capture sufficient structural information to disambiguate similar entities after projection.
    The method rests on the assumption that the projected KG embedding enriches the text representation enough to separate entities like BetaCorp and Beta Corp Spa (Example 2).
  • ad hoc to paper Fine-tuning the baseline on the same textual QA pairs gives a fair comparison with ALIGNed-LLM.
    The ALIGNed model receives an additional projection-layer pretraining stage; the baseline does not. The paper does not control for extra training time or parameter count.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Aligning Knowledge Graphs and Language Models for Factual Accuracy." pith.science (2026). https://pith.science/paper/OZLMPDPZ

@misc{pith2026250713411,
  author       = {Pith},
  title        = {Pith review of: Aligning Knowledge Graphs and Language Models for Factual Accuracy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZLMPDPZ}},
  note         = {Machine review of arXiv:2507.13411}
}
read the original abstract

Large language models like GPT-4, Gemini, and Claude have transformed natural language processing (NLP) tasks such as question answering, dialogue generation, summarization, and so forth; yet their susceptibility to hallucination stands as one of the major challenges. Among numerous approaches to overcome this challenge, integration of Knowledge Graphs (KGs) into language models has emerged as a promising solution as it provides structured, reliable, domain-specific, and up-to-date external information to the language models. In this paper, we introduce ALIGNed-LLM, a simple yet effective approach to improve language models' factuality via a lean strategy to infuse KGs into the latent space of language models inspired by LLaVA where visual and textual information is infused. We use embeddings from a pre-trained Knowledge Graph Embedding (KGE) model, such as TransE, and a trainable projection layer to align entity and text embeddings. This alignment enables the language model to distinguish between similar entities improving factual grounding and reducing hallucination. We tested our approach on three popular questions-answering benchmark datasets alongside language models of varying sizes, showing significant improvement. Furthermore, we applied our approach to a real-world financial use case from a large central bank in Europe, which demands high accuracy and precision, demonstrating a substantial improvement of the LLM answers.

Figures

Figures reproduced from arXiv: 2507.13411 by the authors.

Figure 2
Figure 2. ALIGNed-LLM Pipeline. Pre-trained KG embeddings (left) are used to project text embeddings for subsequent QA (right). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of our approach, showing the workflow of ALIGNed-LLM model. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Impact of Linear vs. Complex Projections across Datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Evaluation metric analysis of ALIGNed Vicuna-13B model on all datasets. method (ALIGNed Mistral-7B) gains around 35.80%, and the ALIGNed TL-1.1B model shows a significant improvement of 10.50% for Wikidata. These results with ALIGNed-LLM emphasize an enhanced structura…
Figure 6
Figure 6. Figure 6: Feature alignment training loss of ALIGNed TL-1.1B model [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 28 canonical work pages

  1. [1]

    Gemini: A family of highly capable multimodal models

    Rohan Anil et al. “Gemini: A family of highly capable multimodal models”. In: arXiv preprint arXiv:2312.11805 1 (2023)

    arXiv 2023

  2. [2]

    Claude: An AI Assistant

    Anthropic. Claude: An AI Assistant . Accessed: 2025-1-

    work page 2025

  3. [3]

    URL: https://www.anthropic.com/product

    2023. URL: https://www.anthropic.com/product

    work page 2023

  4. [4]

    Knowledge-augmented language model verification

    Jinheon Baek, Soyeong Jeong, Minki Kang, et al. “Knowledge-augmented language model verification”. In: arXiv preprint arXiv:2310.12836 (2023)

    arXiv 2023

  5. [5]

    Fine-Tuning Large Enter- prise Language Models via Ontological Reasoning

    Teodoro Baldazzi et al. “Fine-Tuning Large Enter- prise Language Models via Ontological Reasoning”. In: RuleML+RR. V ol. 14244. 2023, pp. 86–94

    work page 2023

  6. [6]

    Softening Ontological Rea- soning with Large Language Models

    Teodoro Baldazzi et al. “Softening Ontological Rea- soning with Large Language Models”. In: RuleML+RR (Companion). V ol. 3816. 2024

    work page 2024

  7. [7]

    Gett-qa: Graph embedding based t2t transformer for knowledge graph question answering

    Debayan Banerjee et al. “Gett-qa: Graph embedding based t2t transformer for knowledge graph question answering”. In: European Semantic Web Conference . Springer. 2023, pp. 279–297

    work page 2023

  8. [8]

    Llms will always hallucinate, and we need to live with this

    Sourav Banerjee, Ayushi Agarwal, and Saloni Singla. “Llms will always hallucinate, and we need to live with this”. In: arXiv preprint arXiv:2409.05746 (2024)

    arXiv 2024

Show all 47 references
  1. [9]

    Besold et al., eds

    Tarek R. Besold et al., eds. Neural-Symbolic Learning and Reasoning . V ol. 14979. NeSy in Lecture Notes in Computer Science. 2024

    2024

  2. [10]

    Translating embeddings for modeling multi-relational data

    Antoine Bordes et al. “Translating embeddings for modeling multi-relational data”. In: Advances in neural information processing systems 26 (2013)

    2013

  3. [11]

    Language Models are Few-Shot Learners

    Tom B. Brown, Benjamin Mann, Nick Ryder, et al. “Language Models are Few-Shot Learners”. In: NeurIPS. 2020

    2020

  4. [12]

    Sparks of artificial general intelligence: Early experiments with gpt-4

    S ´ebastien Bubeck et al. “Sparks of artificial general intelligence: Early experiments with gpt-4”. In: arXiv preprint arXiv:2303.12712 (2023)

    2023 arXiv

  5. [13]

    Knowledge Graph Embedding: A Survey from the Perspective of Representation Spaces. arXiv 2022

    J Cao et al. “Knowledge Graph Embedding: A Survey from the Perspective of Representation Spaces. arXiv 2022”. In: arXiv preprint arXiv:2211.03536 ()

    2022 arXiv

  6. [14]

    REANO: Optimising Retrieval-Augmented Reader Models through Knowledge Graph Generation

    Jinyuan Fang, Zaiqiao Meng, and Craig Macdon- ald. “REANO: Optimising Retrieval-Augmented Reader Models through Knowledge Graph Generation”. In: ACL. 2024, pp. 2094–2112

    2024

  7. [15]

    Mitigating large language model hallucinations via autonomous knowledge graph-based retrofitting

    Xinyan Guan et al. “Mitigating large language model hallucinations via autonomous knowledge graph-based retrofitting”. In: AAAI Conference . V ol. 38. 16. 2024, pp. 18126–18134

    2024

  8. [16]

    Distributed Company Control in Company Shareholding Graphs

    Andrea Gulino et al. “Distributed Company Control in Company Shareholding Graphs”. In: ICDE. 2021, pp. 2637–2648

    2021

  9. [17]

    Knowledgenavigator: Leveraging large language models for enhanced reasoning over knowledge graph

    Tiezheng Guo et al. “Knowledgenavigator: Leveraging large language models for enhanced reasoning over knowledge graph”. In: Complex & Intelligent Systems 10.5 (2024), pp. 7063–7076

    2024

  10. [18]

    A Survey on Knowledge Graphs: Representation, Acquisition, and Applications

    Shaoxiong Ji et al. “A Survey on Knowledge Graphs: Representation, Acquisition, and Applications”. In: IEEE Trans. Neural Networks Learn. Syst. 33.2 (2022), pp. 494–514

    2022

  11. [19]

    Survey of hallucination in natural lan- guage generation

    Ziwei Ji et al. “Survey of hallucination in natural lan- guage generation”. In: ACM Computing Surveys 55.12 (2023), pp. 1–38

    2023

  12. [20]

    Mistral 7B

    Albert Q Jiang et al. “Mistral 7B”. In: arXiv preprint arXiv:2310.06825 (2023)

    2023 arXiv

  13. [21]

    Ef- ficient Knowledge Infusion via KG-LLM Alignment

    Zhouyu Jiang, Ling Zhong, Mengshu Sun, et al. “Ef- ficient Knowledge Infusion via KG-LLM Alignment”. In: arXiv preprint arXiv:2406.03746 (2024)

    2024 arXiv

  14. [22]

    Kagnet: Knowledge-aware graph networks for commonsense reasoning

    Bill Yuchen Lin et al. “Kagnet: Knowledge-aware graph networks for commonsense reasoning”. In: arXiv preprint arXiv:1909.02151 (2019)

    2019 arXiv

  15. [23]

    Visual instruction tuning

    Haotian Liu et al. “Visual instruction tuning”. In: Ad- vances in neural information processing systems 36 (2024)

    2024

  16. [24]

    Chatkbqa: A generate-then-retrieve framework for knowledge base question answering with fine-tuned large language models

    Haoran Luo et al. “Chatkbqa: A generate-then-retrieve framework for knowledge base question answering with fine-tuned large language models”. In: arXiv preprint arXiv:2310.08975 (2023)

    2023 arXiv

  17. [25]

    Knowla: Enhancing parameter- efficient finetuning with knowledgeable adaptation

    Xindi Luo et al. “Knowla: Enhancing parameter- efficient finetuning with knowledgeable adaptation”. In: arXiv preprint arXiv:2403.14950 (2024)

    2024 arXiv

  18. [26]

    Dataset: YAGO3-10 (Yet Another Great Ontology 3-10)

    Farzaneh Mahdisoltani, Joanna Biega, and Fabian Mar- tin Suchanek. Dataset: YAGO3-10 (Yet Another Great Ontology 3-10) . 2024. DOI: 10.57702/g3f0fipm. URL: https://doi.org/10.57702/g3f0fipm

    2024 doi

  19. [27]

    When not to trust language models: Investigating effectiveness and limitations of paramet- ric and non-parametric memories

    Alex Mallen et al. “When not to trust language models: Investigating effectiveness and limitations of paramet- ric and non-parametric memories”. In: arXiv preprint arXiv:2212.10511 7 (2022)

    2022 arXiv

  20. [28]

    Gpt-4 technical report, 2024

    Josh Achiam OpenAI et al. “Gpt-4 technical report, 2024”. In: URL https://arxiv. org/abs/2303.08774 2 (2024), p. 6

    2024 arXiv

  21. [29]

    Unifying large language models and knowledge graphs: A roadmap

    Shirui Pan et al. “Unifying large language models and knowledge graphs: A roadmap”. In: IEEE Transactions on Knowledge and Data Engineering (2024)

    2024

  22. [30]

    Relation-aware language-graph transformer for question answering

    Jinyoung Park et al. “Relation-aware language-graph transformer for question answering”. In: AAAI Confer- ence. 2023, pp. 13457–13464

    2023

  23. [31]

    Check your facts and try again: Improving large language models with external knowl- edge and automated feedback

    Baolin Peng et al. “Check your facts and try again: Improving large language models with external knowl- edge and automated feedback”. In: arXiv preprint arXiv:2302.12813 (2023)

    2023 arXiv

  24. [32]

    Knowledge enhanced con- textual word representations

    Matthew E Peters et al. “Knowledge enhanced con- textual word representations”. In: arXiv preprint arXiv:1909.04164 (2019)

    2019 arXiv

  25. [33]

    Mintaka: A complex, natural, and multilingual dataset for end-to-end question answering

    Priyanka Sen, Alham Fikri Aji, and Amir Saffari. “Mintaka: A complex, natural, and multilingual dataset for end-to-end question answering”. In: arXiv preprint arXiv:2210.01613 (2022)

    2022 arXiv

  26. [34]

    JointLK: Joint reasoning with language models and knowledge graphs for com- monsense question answering

    Yueqing Sun et al. “JointLK: Joint reasoning with language models and knowledge graphs for com- monsense question answering”. In: arXiv preprint arXiv:2112.02732 (2021)

    2021 arXiv

  27. [35]

    Rotate: Knowledge graph embed- ding by relational rotation in complex space

    Zhiqing Sun et al. “Rotate: Knowledge graph embed- ding by relational rotation in complex space”. In: arXiv preprint arXiv:1902.10197 (2019)

    2019 arXiv

  28. [36]

    KG-adapter: Enabling knowledge graph integration in large language models through parameter-efficient fine-tuning

    Shiyu Tian et al. “KG-adapter: Enabling knowledge graph integration in large language models through parameter-efficient fine-tuning”. In: Findings of the Association for Computational Linguistics ACL 2024 . 2024, pp. 3813–3828

    2024

  29. [37]

    Llama: Open and efficient foundation language models

    Hugo Touvron et al. “Llama: Open and efficient foundation language models”. In: arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  30. [38]

    Knowledge graph embedding: A sur- vey of approaches and applications

    Quan Wang et al. “Knowledge graph embedding: A sur- vey of approaches and applications”. In: IEEE transac- tions on knowledge and data engineering 29.12 (2017), pp. 2724–2743

    2017

  31. [39]

    KEPLER: A unified model for knowledge embedding and pre-trained language repre- sentation

    Xiaozhi Wang et al. “KEPLER: A unified model for knowledge embedding and pre-trained language repre- sentation”. In: TACL 9 (2021), pp. 176–194

    2021

  32. [40]

    Resolving knowledge con- flicts in large language models

    Yike Wang et al. “Resolving knowledge con- flicts in large language models”. In: arXiv preprint arXiv:2310.00935 (2023)

    2023 arXiv

  33. [41]

    Thinking with Knowledge Graphs: Enhancing LLM Reason- ing Through Structured Data

    Xue Wu and Kostas Tsioutsiouliklis. “Thinking with Knowledge Graphs: Enhancing LLM Reason- ing Through Structured Data”. In: arXiv preprint arXiv:2412.10654 (2024)

    2024 arXiv

  34. [42]

    QA-GNN: Reasoning with language models and knowledge graphs for question an- swering

    Michihiro Yasunaga et al. “QA-GNN: Reasoning with language models and knowledge graphs for question an- swering”. In: arXiv preprint arXiv:2104.06378 (2021)

    2021 arXiv

  35. [43]

    Tinyllama: An open-source small language model

    Peiyuan Zhang et al. “Tinyllama: An open-source small language model”. In: arXiv preprint arXiv:2401.02385 (2024)

    2024 arXiv

  36. [44]

    ERNIE: Enhanced language representation with informative entities

    Zhengyan Zhang et al. “ERNIE: Enhanced language representation with informative entities”. In: arXiv preprint arXiv:1905.07129 (2019)

    2019 arXiv

  37. [45]

    Fine-tuning Large Language Mod- els for Domain-specific Machine Translation

    Jiawei Zheng et al. “Fine-tuning Large Language Mod- els for Domain-specific Machine Translation”. In: arXiv preprint arXiv:2402.15061 (2024)

    2024 arXiv

  38. [46]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng et al. “Judging llm-as-a-judge with mt-bench and chatbot arena”. In: Advances in Neural Information Processing Systems 36 (2023), pp. 46595– 46623

    2023

  39. [47]

    GraphVite: A High-Performance CPU-GPU Hybrid System for Node Embedding

    Zhaocheng Zhu et al. “GraphVite: A High-Performance CPU-GPU Hybrid System for Node Embedding”. In: The World Wide Web Conference . ACM. 2019, pp. 2494–2504. APPENDIX A. Question-Answering dataset The anonymized company ownership question-answering dataset is built according th...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.