Pith. sign in

REVIEW 5 major objections 5 minor 56 references

Beyond Completion: A Foundation Model for General Knowledge Graph Reasoning

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MERRY claims a single pretrained model fusing graph structure with text beats graph-only baselines on zero-shot link prediction and top LM+KG baselines on CommonsenseQA.

desk verdict MERRY is a sound, code-releasing extension of ULTRA to text-aware KG reasoning, but the headline 'consistently outperforms' claim is not yet statistically grounded and one ablation contradicts it. read the letter →

arxiv 2505.21926 v1 pith:QMPJRFCI submitted 2025-05-28 cs.CL cs.AI

Yin Hua , Zhiqiang Liu , Mingyang Chen , Zheng Fang , Chi Man Wong , Lingxiao Li , Chi Man Vong , Huajun Chen
show 1 more author
Wen Zhang
This is my paper
classification cs.CLcs.AI
keywords knowledgegraphreasoningfoundationmodelinductivelinkpredictionquestionansweringtext-structurefusionconditionalmessagepassingzero-shottransferCommonsenseQA
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 claims that a single pretrained model can handle both knowledge-graph completion and knowledge-graph question answering by combining graph structure with textual descriptions, instead of treating each task with a separate architecture. The authors introduce MERRY, which encodes a query-conditioned structural view and a text-initialized global view of the same graph, fuses them, and scores candidate answers. Across 24 zero-shot inductive completion datasets, MERRY reports higher average MRR and Hits@10 than graph-only foundation models ULTRA and ProLINK, and on CommonsenseQA it reports 74.9% test accuracy, slightly above the LM+KG baseline GreaseLM. A sympathetic reader would take the contribution to be evidence that text and structure are complementary signals that one model can exploit across in-KG and out-of-KG tasks.

What carries the argument

The load-bearing object is the pair of Conditional Message Passing (CMP) channels, where CMP updates a node's representation conditioned on a source node and the query relation rather than pooling neighbors unconditionally. QCMP runs CMP on a relation graph built from four meta-relations (h2h, h2t, t2h, t2t) and then on the entity graph, producing query-specific structural embeddings that are entity- and relation-agnostic. GCMP runs the same CMP machinery with initial embeddings derived from the last token of a frozen LLM (Llama3 8B) over entity and relation descriptions, producing a global semantic channel. A multi-layer perceptron fuses the two channels; DTAF compresses raw text with cross-attention into fixed-length vectors and blends them with the structural channel using learned weights; and a bilinear edge scorer assigns query-relevance weights to edges, which matters for noisy KGQA subgraphs. Training is self-supervised link prediction on three hybrid KGs, with QCMP initialized from ULTRA weights, then all components unfrozen.

What would settle it

Run MERRY on one of the IndER(WK) datasets with all entity descriptions replaced by random tokens: if the MRR advantage over ULTRA does not collapse, the text channel is not doing the work the paper assigns it.

Watch

Extended reading notes

Core claim

MERRY treats knowledge graph reasoning as a single scoring problem: given a query (head, relation, ?) and a graph with text, it produces scores over candidates. Its central claim is that fusing graph structure with text through two complementary message-passing channels—one query-conditioned on structure, one initialized from frozen-LLM text embeddings—lets one pretrained model transfer zero-shot to unseen graphs, including graphs with unseen relations, while also adapting to KGQA with a few examples. Experiments on 24 inductive KGC datasets report average MRR 0.445 and Hits@10 0.626, above ULTRA and ProLINK; on CommonsenseQA test, accuracy is 74.9%, above GreaseLM's 74.2%. The paper therefore claims that a single framework can be competitive in both in-KG and out-of-KG reasoning.

Load-bearing premise

The model assumes every entity and relation has a usable textual description and that the last-token embedding from the frozen LLM captures enough of that text; the paper's own limitations section notes that some datasets have missing textual fields.

Editorial extensions

If this is right

  • If the central claim holds, KGQA can reuse the same pretrained graph encoder as completion, needing only a few-shot adaptation for the answer relation.
  • Zero-shot inductive KGC on graphs with unseen relations improves when entity and relation text is available, so graph-only methods leave signal on the table.
  • The same query format (head, relation, ?) covers both completion and QA, making task-specific encoders unnecessary.
  • Because text features are precomputed offline and the LM stays frozen, deployment cost on large graphs stays near GNN-level rather than LLM-level.
  • Swapping the frozen LLM backbone does not require retraining the graph modules, since the text channel is parameter-free at inference time.

Reading between the lines

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

  • Editorial inference: the paper's own limitation about missing textual fields implies that enriching or generating descriptions for sparse KGs should make the text channel even more valuable than the reported average gains suggest.
  • Editorial inference: the learned fusion weights in DTAF could be read as a per-task measure of how much a dataset relies on text versus structure; the ablation pattern across KGC and CSQA supports that reading.
  • Editorial inference: the edge-scoring mechanism, currently bilinear on text features, may extend to other graph tasks such as retrieval or recommendation where edge relevance is query-dependent.
  • Editorial inference: a direct test of the text channel's quality would be to keep graph weights fixed, swap the frozen LLM for a stronger one, and measure zero-shot KGC gains, isolating text encoding quality from structural learning.
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 / 5 minor

Summary. The paper introduces MERRY, a foundation model for knowledge graph reasoning that integrates structural and textual information through query-conditioned message passing (QCMP and GCMP), a dynamic text-adaptive fusion module (DTAF), and a task-adaptive edge scoring mechanism. The model is pre-trained on WN18RR, FB15k237, and CoDEx-Medium and then evaluated zero-shot on 27 inductive KGC datasets and (after few-shot adaptation) on CommonsenseQA. The paper reports average MRR 0.445 and Hits@10 0.626 over 24 headline KGC datasets, and 74.9% test accuracy on CSQA, claiming consistent improvements over ULTRA, ProLINK, and KGQA baselines such as GreaseLM.

Significance. If the reported results hold, MERRY would be a useful step toward a single model that handles both in-KG and out-of-KG reasoning tasks, with a clean architectural separation between query-specific structural encoding and global semantic encoding. The paper provides a detailed description of the architecture, releases code, and evaluates on a broad set of datasets, which are assets. However, the central empirical claim is not yet statistically grounded: the reported margins are small, no uncertainty is quantified, and key ablations are confounded or show a negative marginal effect. The current evidence is therefore insufficient for the strength of the conclusions, though the underlying approach is defensible and worth revising.

major comments (5)
  1. [§5.4, Table 1] The claim that MERRY 'consistently outperforms' baselines is not supported by the reported numbers: the average MRR advantage over ULTRA and ProLINK is 0.012 (0.445 vs 0.433), the per-dataset best count is 12/24, and no standard deviations, confidence intervals, or significance tests are provided for any result. Small differences of this magnitude could be run-to-run noise, especially on datasets with only a few thousand test triples. Please report multiple seeds with variance and a paired test, and soften the 'consistently outperforms' claim accordingly.
  2. [§4.6 and Table 2] The KGQA comparison is not controlled: MERRY is fine-tuned with only three few-shot examples (Section 4.6), while the baselines such as GreaseLM and QA-GNN are trained on the full in-house training split. If MERRY indeed uses only three examples, the comparison conflates model quality with data efficiency; if it uses the full training set, the methods text contradicts the description. Please clarify the exact training protocol and either compare against baselines under the same data regime or explicitly frame the result as a few-shot comparison.
  3. [Table 3] The edge-scoring ablation is confounded: the row labeled 'w/o Edge Scoring' also removes DTAF, so the drop from 74.9 to 70.7 cannot be attributed to edge scoring alone. In addition, removing DTAF while keeping edge scoring improves IHtest accuracy from 74.9 to 75.0, which contradicts the text's claim that 'DTAF significantly impacts KGQA performance.' Please rerun one-at-a-time ablations with multiple seeds and report the marginal contribution of each module.
  4. [§4.3 and Limitations] The paper states that entity and relation text is encoded by the LLM and used as X_e in Eq. (9), but the Limitations section admits that some datasets have missing textual fields for certain entities. The paper does not specify how missing descriptions are handled (e.g., placeholder embeddings, dropping the node, or imputation). This matters because the GCMP channel becomes vacuous for those nodes, and the KGC results on such datasets are hard to interpret. Please specify the fallback and report which datasets have missing fields and the impact on results.
  5. [§5.3 and Table 6] The zero-shot generalization claim is weakened by source overlap: the model is pre-trained on WN18RR, FB15k237, and CoDEx-Medium, and the IndE(WN), IndE(FB), and IndER(FB) evaluation sets are derived from WN18RR and FB15k237. The reported gains on these benchmarks may partly reflect in-distribution text and relation statistics rather than generalization to truly unseen KGs. Please report a version of the average metrics computed on benchmarks whose source KGs were not in pre-training, or otherwise analyze the sensitivity of the conclusion to this overlap.
minor comments (5)
  1. [References] The reference 'ChatGPT and Richard L. Barnes. 2023. AI sarcasm detection: Insult your AI without offending it. RFC 9405' is unrelated to the cited context and appears to be an erroneous entry; please remove or replace it. Also, Raffel et al. is cited as 2023 but the T5 paper is from 2020.
  2. [§5.4] The text says '27 inductive link prediction KG datasets, categorized into 7 benchmarks,' but Table 1 presents 6 benchmarks and 24 datasets; please align the counts and clarify why ILPC-small, ILPC-large, and NL-0 are excluded from the main table.
  3. [Eq. (17)] The variable x_q is used in Eq. (17) but is not defined before the equation; please define it and clarify how the query text is tokenized and encoded.
  4. [Figures 2 and 3] Figures 2 and 3 lack error bars and axis labels; please add them so the ablation and hyperparameter results can be properly assessed.
  5. [Table 1] The 'MERRY_PNA' variant is not described in the methodology section; please define PNA and explain how it differs from the main MERRY model.

Circularity Check

1 steps flagged · score 6.0 of 10

Zero-shot KGC on 12 of 24 headline datasets is in-sample: MERRY is pre-trained on WN18RR and FB15k237 full KGs, the same sources as the IndE and IndER(FB) inductive test graphs.

  1. fitted input called prediction [Section 5.1 (Datasets and Metrics), Section 5.3 (Implementation & Training details), Appendix C (Datasets), Eq. 18]
    "We pre-train MERRY on three hybrid knowledge graph datasets: WN18RR, CoDEx-Medium, and FB15k237... This category includes 12 datasets from (Teru et al., 2020): WN18RR (WN), FB15k237 (FB), and NELL-995 (NL)... This category comprises 13 graphs from (Lee et al., 2023): FB15k237 (FB) and Wikidata68K (WK)..."

    The IndE(WN), IndE(FB), and IndER(FB) test graphs are inductive splits of WN18RR and FB15k237, the same full KGs used for pre-training. The pre-training objective (Eq. 18) is link prediction over triples of those full KGs, so the exact triples scored as 'zero-shot' test queries in those benchmarks were already seen and optimized during pre-training. Reporting MRR and Hits@10 on these subsets as evidence of zero-shot generalization is therefore partially a re-measurement of the training objective rather than an out-of-sample prediction. This affects 12 of the 24 headline datasets in Table 1 and inflates the total average, although the remaining 12 datasets (NELL, Wikidata68K, ILPC, and CSQA) are not contaminated.

full rationale

MERRY is an empirical systems paper, not a mathematical derivation, so the usual circularity modes (self-definition, uniqueness imported from authors, ansatz smuggled via citation) do not apply. The main circularity-adjacent defect is pre-training/test overlap: 12 of the 24 headline zero-shot KGC datasets are inductive splits of WN18RR and FB15k237, the same full KGs on which MERRY is pre-trained with the link-prediction loss of Eq. 18, making those 'zero-shot' numbers partially in-sample. The remaining benchmarks (NELL, Wikidata68K, ILPC) and CSQA are not contaminated and MERRY still leads there, so the paper has independent content. ULTRA is also pre-trained on the same three graphs, so the head-to-head comparison is less affected; the generalization claim is what suffers. The DTAF ablation is internally inconsistent (removing DTAF raises IHtest accuracy from 74.9 to 75.0 in Table 3 while the text credits DTAF for KGQA gains), and no significance testing is reported for the small aggregate margins; those are correctness risks, not circularity. The self-citations (UniHR, OntoTune) are contextual and not load-bearing. Overall, the paper's central zero-shot claim is partially circular because a substantial block of its evaluation data is drawn from the pre-training corpora.

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

MERRY is an empirical deep learning system, so its free parameters are the learned weights and hand-chosen hyperparameters of the architecture. The core axioms are inherited from prior work: CMP from Huang et al., relation graph from ULTRA, and the usefulness of frozen LLM text embeddings. The paper itself acknowledges the fragility of the text-description assumption in the Limitations section. The invented entities are purely task-formulation scaffolding for KGQA, not new physical or mathematical objects.

free parameters (6)
  • alpha (text-structure fusion weight for relations) = learned
    Equation 15: Rf = alpha * Xr + (1-alpha) * RCMP. Learned during training to balance textual and structural features.
  • beta (text-structure fusion weight for entities) = learned
    Equation 16: Hf = beta * Xe + (1-beta) * HCMP. Learned to balance modalities per task.
  • Qtoken (trainable query tokens in DTAF cross-attention) = learned (dimension d x k)
    Equations 12-14: Qtoken is a trainable parameter matrix used to pool token-level text into fixed-length embeddings.
  • W (bilinear coefficient in edge scoring) = learned (3d x d)
    Equation 17: eta = Norm([xh||xr||xt]^T W xq). Learned for query-conditional edge relevance.
  • Rmeta and Rhat_meta (meta-relation embedding matrices) = learned (4 x d each)
    Equations 5 and 8: these matrices encode the four fixed meta-relations (h2h, h2t, t2h, t2t) for structural and textual perspectives.
  • GCMP layer count and k (number of DTAF query tokens) = 3 layers, k tuned
    Section 5.7 studies GCMP depth; k is described as a tunable hyperparameter. These are chosen by hand rather than derived.
assumptions (5)
  • domain assumption Conditional Message Passing (CMP), as defined in Huang et al. (2023), is an effective graph encoder for triple-level interactions.
    The entire architecture is built from CMP units (Section 4.1), but the paper does not prove or verify the effectiveness of CMP; it relies on the cited publication.
  • domain assumption The relation graph lifting defined in ULTRA (Galkin et al., 2024) produces dataset-agnostic relation representations that transfer to unseen entities and relations.
    Section 4.2 follows ULTRA's construction and the paper's own results show that transfer works only partially; the assumption is foundational to the zero-shot KGC claim.
  • domain assumption The last-token hidden state of a frozen Llama3 8B provides useful semantic embeddings for entities and relations.
    Section 4.3 adopts a parameter-free strategy of extracting the last token from LLM output. There is no analysis validating that this vector captures the required semantics beyond the reported end-task accuracy.
  • domain assumption Every entity and relation in the evaluation datasets has a usable textual description.
    Section 4.3 assumes textual descriptions exist for all nodes and relations. The Limitations section explicitly notes that some datasets have missing textual fields, so the assumption is known to be violated.
  • domain assumption In KGQA, inserting an artificial relation REL_the_answer_is and using a few-shot Sentence-BERT retrieval strategy is sufficient to bridge the gap between link prediction and question answering.
    Section 4.6 introduces this adaptation. The zero-shot KGQA performance is poor and improves only after few-shot fine-tuning, indicating the assumption is fragile.
invented entities (3)
  • REL_the_answer_is
    purpose: An artificial relation connecting question nodes to answer nodes, unifying KGQA with the KGC query format (h, r, ?).
    Introduced in Section 3 and Section 4.6. It is a synthetic relation not present in the original knowledge graphs and has no independent falsifiable handle outside the model's own task formulation.
  • question-node
    purpose: A graph node representing the input question, connected to topic entities via a new relation.
    Introduced in Section 4.6. A task-specific construct that has no existence in the underlying KG and carries only the question text.
  • answer-node
    purpose: A graph node representing each candidate answer option, connected to entities mentioned in the option.
    Introduced in Section 4.6. A synthetic node created to make answer selection a link prediction task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Completion: A Foundation Model for General Knowledge Graph Reasoning." pith.science (2026). https://pith.science/paper/QMPJRFCI

@misc{pith2026250521926,
  author       = {Pith},
  title        = {Pith review of: Beyond Completion: A Foundation Model for General Knowledge Graph Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMPJRFCI}},
  note         = {Machine review of arXiv:2505.21926}
}
read the original abstract

In natural language processing (NLP) and computer vision (CV), the successful application of foundation models across diverse tasks has demonstrated their remarkable potential. However, despite the rich structural and textual information embedded in knowledge graphs (KGs), existing research of foundation model for KG has primarily focused on their structural aspects, with most efforts restricted to in-KG tasks (e.g., knowledge graph completion, KGC). This limitation has hindered progress in addressing more challenging out-of-KG tasks. In this paper, we introduce MERRY, a foundation model for general knowledge graph reasoning, and investigate its performance across two task categories: in-KG reasoning tasks (e.g., KGC) and out-of-KG tasks (e.g., KG question answering, KGQA). We not only utilize the structural information, but also the textual information in KGs. Specifically, we propose a multi-perspective Conditional Message Passing (CMP) encoding architecture to bridge the gap between textual and structural modalities, enabling their seamless integration. Additionally, we introduce a dynamic residual fusion module to selectively retain relevant textual information and a flexible edge scoring mechanism to adapt to diverse downstream tasks. Comprehensive evaluations on 28 datasets demonstrate that MERRY outperforms existing baselines in most scenarios, showcasing strong reasoning capabilities within KGs and excellent generalization to out-of-KG tasks such as KGQA.

Figures

Figures reproduced from arXiv: 2505.21926 by the authors.

Figure 1
Figure 1. Overview of the MERRY Framework. (A) All tasks, including KGC and KGQA, are unified under a [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Performance of different GCMP layers in KGC and different numbers of shots in KGQA. We investigated the impact of GCMP layers on zero-shot KGC tasks and assessed the role of few￾shot learning in KGQA. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 2
Figure 2. Ablation study results. The results demonstrate a significant perfor￾mance drop in the "w/o GCMP" variant, highlight￾ing its critical role in bridging the gap between tex￾tual and structural modalities for better integration. In contrast, the "w/o DTAF" variant shows a slight performance decline, indicating that while original text features aid KGC, DTAF primarily enhances the understanding of structural information… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 22 canonical work pages

  1. [1]

    Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. Advances in neural information processing systems, 26

    2013

  2. [2]

    ChatGPT and Richard L. Barnes. 2023. https://doi.org/10.17487/RFC9405 AI sarcasm detection: Insult your AI without offending it . RFC , 9405:1--5

    doi:10.17487/rfc9405 2023

  3. [3]

    Jiajun Chen, Huarui He, Feng Wu, and Jie Wang. 2021. https://arxiv.org/abs/2103.03642 Topology-aware correlations between relations for inductive link prediction in knowledge graphs . Preprint, arXiv:2103.03642

    work page Pith review arXiv 2021

  4. [4]

    Mingyang Chen, Wen Zhang, Zhen Yao, Xiangnan Chen, Mengxiao Ding, Fei Huang, and Huajun Chen. 2022. https://doi.org/10.24963/ijcai.2022/273 Meta-learning based knowledge extrapolation for knowledge graphs in the federated setting . In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22 , pages 1966--1972. In...

    work page doi:10.24963/ijcai.2022/273 2022

  5. [5]

    Gabriele Corso, Luca Cavalleri, Dominique Beaini, Pietro Liò, and Petar Veličković. 2020. https://arxiv.org/abs/2004.05718 Principal neighbourhood aggregation for graph nets . Preprint, arXiv:2004.05718

    arXiv 2020

  6. [6]

    Yuanning Cui, Yuxin Wang, Zequn Sun, Wenqiang Liu, Yiqiao Jiang, Kexin Han, and Wei Hu. 2022. https://arxiv.org/abs/2208.10378 Inductive knowledge graph reasoning for multi-batch emerging entities . Preprint, arXiv:2208.10378

    work page Pith review arXiv 2022

  7. [7]

    Daniel Daza, Michael Cochez, and Paul Groth. 2021. https://doi.org/10.1145/3442381.3450141 Inductive entity representations from text via link prediction . In Proceedings of the Web Conference 2021, WWW ’21, page 798–808. ACM

    arXiv 2021

  8. [8]

    Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. https://arxiv.org/abs/1707.01476 Convolutional 2d knowledge graph embeddings . Preprint, arXiv:1707.01476

    arXiv 2018

Show all 56 references
  1. [9]

    Yanlin Feng, Xinyue Chen, Bill Yuchen Lin, Peifeng Wang, Jun Yan, and Xiang Ren. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.99 Scalable multi-hop relational reasoning for knowledge-aware question answering . In Proceedings of the 2020 Conference on Empirical Methods in ...

    2020 doi

  2. [10]

    Mikhail Galkin, Max Berrendorf, and Charles Tapley Hoyt. 2022 a . https://doi.org/10.48550/ARXIV.2203.01520 An open challenge for inductive link prediction on knowledge graphs . CoRR, abs/2203.01520

    2022 doi arXiv

  3. [11]

    Hamilton

    Mikhail Galkin, Etienne Denis, Jiapeng Wu, and William L. Hamilton. 2022 b . https://arxiv.org/abs/2106.12144 Nodepiece: Compositional and parameter-efficient representations of large knowledge graphs . Preprint, arXiv:2106.12144

    2022 arXiv

  4. [12]

    Mikhail Galkin, Xinyu Yuan, Hesham Mostafa, Jian Tang, and Zhaocheng Zhu. 2024. https://openreview.net/forum?id=jVEoydFOl9 Towards foundation models for knowledge graph reasoning . In The Twelfth International Conference on Learning Representations

    2024

  5. [13]

    Pan, Mingyang Chen, Song Jiang, Wen Zhang, and Huajun Chen

    Yuxia Geng, Jiaoyan Chen, Jeff Z. Pan, Mingyang Chen, Song Jiang, Wen Zhang, and Huajun Chen. 2022. https://arxiv.org/abs/2210.03994 Relational message passing for fully inductive knowledge graph completion . Preprint, arXiv:2210.03994

    2022 arXiv

  6. [14]

    Genet Asefa Gesese, Harald Sack, and Mehwish Alam. 2023. https://doi.org/10.1145/3579051.3579066 Raild: Towards leveraging relation features for inductive link prediction in knowledge graphs . In Proceedings of the 11th International Joint Conference on Knowledge Graphs, IJCKG...

    2023

  7. [15]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Art...

    2024 arXiv

  8. [16]

    Qingyu Guo, Fuzhen Zhuang, Chuan Qin, Hengshu Zhu, Xing Xie, Hui Xiong, and Qing He. 2020. https://arxiv.org/abs/2003.00911 A survey on knowledge graph-based recommender systems . Preprint, arXiv:2003.00911

    2020 arXiv

  9. [17]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 1025–1035, Red Hook, NY, USA. Curran Associates Inc

    2017

  10. [18]

    Xingyue Huang, Miguel Romero Orth, İsmail İlkan Ceylan, and Pablo Barceló. 2023. https://openreview.net/forum?id=7hLlZNrkt5 A theory of link prediction via relational weisfeiler-leman on knowledge graphs . In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  11. [19]

    Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and Philip S. Yu. 2022. https://doi.org/10.1109/TNNLS.2021.3070843 A survey on knowledge graphs: Representation, acquisition, and applications . IEEE Trans. Neural Networks Learn. Syst. , 33(2):494--514

    2022

  12. [20]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. https://openreview.net/forum?id=SJU4ayYgl Semi-supervised classification with graph convolutional networks . In International Conference on Learning Representations

    2017

  13. [21]

    Jaejun Lee, Chanyoung Chung, and Joyce Jiyoung Whang. 2023. https://arxiv.org/abs/2305.19987 Ingram: Inductive knowledge graph embedding via relation graphs . Preprint, arXiv:2305.19987

    2023 arXiv

  14. [22]

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. 2024. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895

    2024 arXiv

  15. [23]

    Bill Yuchen Lin, Xinyue Chen, Jamin Chen, and Xiang Ren. 2019. https://doi.org/10.18653/v1/D19-1282 K ag N et: Knowledge-aware graph networks for commonsense reasoning . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Inter...

    2019 doi

  16. [24]

    Ye Liu, Kai Zhang, Zhenya Huang, Kehang Wang, Yanghai Zhang, Qi Liu, and Enhong Chen. 2023. https://doi.org/10.18653/v1/2023.findings-acl.358 Enhancing hierarchical text classification through knowledge graph integration . In Findings of the Association for Computational Lingu...

    2023 doi

  17. [25]

    Zhiqiang Liu, Mingyang Chen, Yin Hua, Zhuo Chen, Ziqi Liu, Lei Liang, Huajun Chen, and Wen Zhang. 2024. Unihr: Hierarchical representation learning for unified knowledge graph link prediction. arXiv preprint arXiv:2411.07019

    2024

  18. [26]

    Zhiqiang Liu, Chengtao Gan, Junjie Wang, Yichi Zhang, Zhongpu Bo, Mengshu Sun, Huajun Chen, and Wen Zhang. 2025. https://doi.org/10.1145/3696410.3714816 Ontotune: Ontology-driven self-training for aligning large language models . In Proceedings of the ACM on Web Conference 202...

    2025

  19. [27]

    Shangwen Lv, Daya Guo, Jingjing Xu, Duyu Tang, Nan Duan, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, and Songlin Hu. 2020. https://doi.org/10.1609/aaai.v34i05.6364 Graph-based reasoning over heterogeneous external knowledge for commonsense question answering . Proceeding...

    2020 doi

  20. [28]

    Elan Markowitz, Keshav Balasubramanian, Mehrnoosh Mirtaheri, Murali Annavaram, Aram Galstyan, and Greg Ver Steeg. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.46 S t ATIK : Structure and text for inductive knowledge graph completion . In Findings of the Association fo...

    2022 doi

  21. [29]

    Todor Mihaylov and Anette Frank. 2018. https://doi.org/10.18653/v1/P18-1076 Knowledgeable reader: Enhancing cloze-style reading comprehension with external commonsense knowledge . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volum...

    2018 doi

  22. [30]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2023. https://arxiv.org/abs/1910.10683 Exploring the limits of transfer learning with a unified text-to-text transformer . Preprint, arXiv:1910.10683

    2023 arXiv

  23. [31]

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R \"a dle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Doll \'a r, and Christoph F...

    2024 arXiv

  24. [32]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . Preprint, arXiv:1908.10084

    2019 arXiv

  25. [33]

    Tara Safavi and Danai Koutra. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.669 C o DE x: A C omprehensive K nowledge G raph C ompletion B enchmark . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8328--8350, Online...

    2020 doi

  26. [34]

    Adam Santoro, David Raposo, David G. T. Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia, and Timothy Lillicrap. 2017. https://arxiv.org/abs/1706.01427 A simple neural network module for relational reasoning . Preprint, arXiv:1706.01427

    2017 arXiv

  27. [35]

    Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling

    Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2017. https://arxiv.org/abs/1703.06103 Modeling relational data with graph convolutional networks . Preprint, arXiv:1703.06103

    2017 arXiv

  28. [36]

    Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. 2019. https://arxiv.org/abs/1902.10197 Rotate: Knowledge graph embedding by relational rotation in complex space . Preprint, arXiv:1902.10197

    2019 arXiv

  29. [37]

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. https://arxiv.org/abs/1811.00937 Commonsenseqa: A question answering challenge targeting commonsense knowledge . Preprint, arXiv:1811.00937

    2019 arXiv

  30. [38]

    Teru, Etienne Denis, and William L

    Komal K. Teru, Etienne Denis, and William L. Hamilton. 2020. Inductive relation prediction by subgraph reasoning. arXiv: Learning

    2020

  31. [39]

    Kristina Toutanova and Danqi Chen. 2015. https://doi.org/10.18653/v1/W15-4007 Observed versus latent features for knowledge base and text inference . In Proceedings of the 3rd Workshop on Continuous Vector Space Models and their Compositionality, pages 57--66, Beijing, China. ...

    2015 doi

  32. [40]

    Shikhar Vashishth, Soumya Sanyal, Vikram Nitin, and Partha Talukdar. 2020. https://arxiv.org/abs/1911.03082 Composition-based multi-relational graph convolutional networks . Preprint, arXiv:1911.03082

    2020 arXiv

  33. [41]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. https://arxiv.org/abs/1710.10903 Graph attention networks . Preprint, arXiv:1710.10903

    2018 arXiv

  34. [42]

    Bo Wang, Tao Shen, Guodong Long, Tianyi Zhou, Ying Wang, and Yi Chang. 2021. https://doi.org/10.1145/3442381.3450043 Structure-augmented text representation learning for efficient knowledge graph completion . In Proceedings of the Web Conference 2021, WWW ’21, page 1737–1748. ACM

    2021

  35. [43]

    Kai Wang, Yuwei Xu, Zhiyong Wu, and Siqiang Luo. 2024. https://doi.org/10.18653/v1/2024.findings-acl.224 LLM as prompter: Low-resource inductive reasoning on arbitrary knowledge graphs . In Findings of the Association for Computational Linguistics: ACL 2024, pages 3742--3759, ...

    2024 doi

  36. [44]

    Liang Wang, Wei Zhao, Zhuoyu Wei, and Jingming Liu. 2022. https://aclanthology.org/2022.acl-long.295 S im KGC : Simple contrastive knowledge graph completion with pre-trained language models . In Proceedings of the 60th Annual Meeting of the Association for Computational Lingu...

    2022

  37. [45]

    Xiaoyan Wang, Pavan Kapanipathi, Ryan Musa, Mo Yu, Kartik Talamadupula, Ibrahim Abdelaziz, Maria Chang, Achille Fokoue, Bassem Makni, Nicholas Mattei, and Michael Witbrock. 2018. https://arxiv.org/abs/1809.05724 Improving natural language inference using external knowledge in ...

    2018 arXiv

  38. [46]

    Xiaoyan Wang, Pavan Kapanipathi, Ryan Musa, Mo Yu, Kartik Talamadupula, Ibrahim Abdelaziz, Maria Chang, Achille Fokoue, Bassem Makni, Nicholas Mattei, and Michael Witbrock. 2019. https://doi.org/10.1609/aaai.v33i01.33017208 Improving natural language inference using external k...

    2019 doi

  39. [47]

    Wenhan Xiong, Thien Hoang, and William Yang Wang. 2018. https://arxiv.org/abs/1707.06690 Deeppath: A reinforcement learning method for knowledge graph reasoning . Preprint, arXiv:1707.06690

    2018 arXiv

  40. [48]

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. 2024. https://doi.org/10.1145/3626772.3661370 Retrieval-augmented generation with knowledge graphs for customer service question answering . In Proceedings of the 47th Inter...

    2024

  41. [49]

    An Yang, Quan Wang, Jing Liu, Kai Liu, Yajuan Lyu, Hua Wu, Qiaoqiao She, and Sujian Li. 2019. https://doi.org/10.18653/v1/P19-1226 Enhancing pre-trained language representations with rich knowledge for machine reading comprehension . In Proceedings of the 57th Annual Meeting o...

    2019 doi

  42. [50]

    Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut, Percy Liang, and Jure Leskovec. 2021. Qa-gnn: Reasoning with language models and knowledge graphs for question answering. In North American Chapter of the Association for Computational Linguistics (NAACL)

    2021

  43. [51]

    Xikun Zhang, Antoine Bosselut, Michihiro Yasunaga, Hongyu Ren, Percy Liang, Christopher D Manning, and Jure Leskovec. 2021. Greaselm: Graph reasoning enhanced language models. In International Conference on Learning Representations

    2021

  44. [52]

    Yongqi Zhang and Quanming Yao. 2022. Knowledge graph reasoning with relational digraph. In Proceedings of the ACM Web Conference 2022, pages 912--924

    2022

  45. [53]

    Zhaocheng Zhu, Xinyu Yuan, Mikhail Galkin, Sophie Xhonneux, Ming Zhang, Maxime Gazeau, and Jian Tang. 2023. https://arxiv.org/abs/2206.04798 A*net: A scalable path-based reasoning approach for knowledge graphs . Preprint, arXiv:2206.04798

    2023 arXiv

  46. [54]

    Zhaocheng Zhu, Zuobai Zhang, Louis-Pascal Xhonneux, and Jian Tang. 2021. Neural bellman-ford networks: A general graph neural network framework for link prediction. Advances in Neural Information Processing Systems, 34

    2021

  47. [55]

    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...

  48. [56]

    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 gl...

Pith tools

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