Pith. sign in

REVIEW 4 major objections 5 minor 48 references

Large Language Model-Enhanced Symbolic Reasoning for Knowledge Base Completion

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read LeSR shows that LLM-proposed, knowledge-base-grounded logic rules, reweighted by a trainable rule reasoner and blended with an embedding model, achieve state-of-the-art or competitive accuracy on five KBC benchmarks while keeping rules…

desk verdict A promising LLM-rule-mining pipeline undermined by an invalid loss function and some unreliable empirical claims. read the letter →

arxiv 2501.01246 v1 pith:RTY4KAOH submitted 2025-01-02 cs.CL

Qiyuan He , Jianfei Yu , Wenya Wang This is my paper · ORCID
classification cs.CL
keywords knowledgebasecompletionrule-basedreasoninglargelanguagemodelssymboliclinkpredictioninterpretablerulesruleminingembedding-basedinference
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 LeSR, a framework that uses a large language model to propose candidate logic rules from sampled knowledge-base subgraphs, then trains a rule reasoner to reweight those rules against the observed KB and combines them with an embedding model for knowledge base completion. The claim is that LLM-proposed, KB-grounded rules can match or beat standard KBC methods while remaining human-readable. The authors report state-of-the-art or competitive results across five datasets, with the largest gains on larger, realistic benchmarks such as FB15K-237, WD15K, and CN100. If correct, this offers a path to KBC that is both accurate and interpretable, avoiding the opacity of pure embedding methods and the hallucination of using LLMs directly to answer queries.

What carries the argument

The Rule Reasoner's grounding matrices are the load-bearing object. Each proposed rule $\varphi$ is represented by a body matrix C_i and a head-satisfaction matrix A_i, computed by matrix multiplications and transposes over relation adjacency matrices for categorized rule structures (0th-, 1st-, and 2nd-order paths). A rule scores a candidate triple by s_i(h,t) = A_i(h,t) if the head is found, -C_i(h,t) if the body fires but the head is absent, and 0 otherwise. These per-rule scores are combined with learnable weights w_i (softmax-normalized) and blended with RotatE via a weight $\alpha$, trained with a log-likelihood loss that masks zero contributions. This machinery lets the model quantify which LLM-proposed rules are actually supported by the KB and down-weight hallucinated ones.

What would settle it

Take a KB such as WD15K, hold out a random sample of true triples so they are absent from the training KB, learn the Rule Reasoner weights, and check whether rules whose bodies imply those held-out triples receive negative weights; if they do, the closed-world scoring demonstrably penalizes valid but missing facts.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hybrid reasoner can get the best of both worlds: an LLM proposes diverse, relation-specific IF-THEN rules from subgraph context, and a trainable Rule Reasoner grounds those rules in the knowledge base by counting how often each rule body and head co-occur, scoring each rule, and learning a significance weight for it. The rule weights are optimized together with an embedding score (RotatE) via a log-likelihood objective. On FB15K-237, WD15K, and CN100 LeSR outperforms all baselines across MRR, Hit@1, Hit@3, and Hit@10; on the smaller UMLS and WN18RR it is among the top three. The same pipeline also yields rules with a higher Rule Quality Index than AnyBURL, NeuralLP, and RNNLogic on WD15K, meaning a larger fraction of learned rules are interpretable and reliable.

Load-bearing premise

The rule weights are learned by treating every triple absent from the training KB as false, even though the paper argues elsewhere that knowledge bases are incomplete and absence should mean uncertainty, not falsehood.

Editorial extensions

If this is right

  • Removing the Rule Reasoner's learned weights and treating all LLM-proposed rules equally degrades MRR and Hit@K on all five datasets, so the learned significance scores are what make LLM proposals reliable.
  • Using GPT-4 as the rule proposer produces far more rules than GPT-3.5, but no consistent accuracy gain, because the extra rules tend to be tied to the specific subgraphs in the prompt rather than general KB patterns.
  • On WD15K, LeSR's Rule Quality Index (51.03 for the GPT-4 variant) is about three times AnyBURL's (8.70) and RNNLogic's (17.02), while keeping a much smaller rule set than AnyBURL, showing the rules are not just accurate but interpretable.

Reading between the lines

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

  • The paper evaluates only closed-world standard benchmarks; a natural stress test would be to apply LeSR to a KB where a known fraction of true facts is withheld, and measure whether the Rule Reasoner's negative scoring of absent triples depresses the weights of exactly those rules that would recover the withheld facts.
  • The rule-structure taxonomy is finite (0th, 1st, and 2nd order, with eight 2nd-order shapes), so the paper's evidence does not tell whether longer-chain rules proposed by the LLM would help or hurt; an extension would be to let the Rule Reasoner accept arbitrary-length rules and compare coverage.
  • The relation-name mapping via sentence-transformer similarity is a potential bottleneck: if the LLM proposes a semantically valid but lexically distant relation, the mapping may fail and the rule is discarded before the Reasoner sees it. Testing with different embedding-based mappers would quantify this.
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 / 5 minor

Summary. The manuscript proposes LeSR, a Knowledge Base Completion framework that combines LLM-proposed symbolic rules with a trainable rule reasoner and a RotatE embedding model. The Subgraph Extractor samples relation-centered subgraphs; the LLM Proposer converts them into candidate logic rules; the Rule Reasoner grounds these rules against the KB, learns per-rule significance weights, and mixes rule scores with an embedding score for inference. Experiments are reported on UMLs, WN18RR, FB15K-237, WD15K, and CN100, comparing against AnyBURL, RotatE, NeuralLP, DRUM, RNNLogic, RGCN, and direct LLM inference. The paper claims competitive or state-of-the-art performance and higher rule interpretability on WD15K, with ablations showing the benefit of learning rule weights.

Significance. The idea of using LLMs to propose diverse, relation-specific rules and then grounding them with a trainable symbolic reasoner is timely and potentially useful, especially because the framework produces human-readable rules alongside a neural scorer. The WD15K-based Rule Clarity Score and Rule Quality Index are a notable attempt to quantify rule interpretability, and the ablations on weight learning are informative. If the training objective and the reported baseline numbers are corrected, the framework could be a useful contribution to KBC; as written, however, several load-bearing technical points prevent the results from being fully attributable to the described algorithm.

major comments (4)
  1. [Section 3.4] The training objective is under-specified and, as written, can be undefined. The scoring function sets s_i(h,t) = -C_i(h,t) when A_i(h,t)=0 and C_i(h,t)>0, and f_phi_i is defined as s_i(h,t'); RotatE's score f_emb is typically non-positive. Since w_i and w_emb are made positive by softmax, the argument of the logarithm in L1 = -log(alpha * sum_i w_i f_phi_i + (1-alpha) * w_emb * f_emb) can be non-positive. The statement that zero s_i values are masked removes zeros, not negative values, so it does not fix the problem. The text also does not constrain alpha or clarify how L0 relates to L1. Because the Rule Reasoner's learned weights are the mechanism claimed to suppress hallucinated LLM rules, the actual objective used in the experiments must be reported (e.g., sigmoid-normalized scores, clipping, or a ranking loss); otherwise Tables 2-7 cannot be attributed to the described algorithm.
  2. [Section 3.4 and Section 2] The rule-evaluation signal implements a closed-world assumption that conflicts with the paper's own critique of embedding methods. By penalizing every body-satisfying but head-absent pair with -C_i(h,t), the reasoner treats all absent triples as false. This is exactly the 'closed world assumption' that Section 2 criticizes in embedding-based methods. Because the learned rule weights are driven by this negative signal, the method inherits the limitation the paper uses to motivate it. The authors should either adopt an open-world training signal (e.g., filtered negatives or an explicit uncertainty model) or justify why a local closed-world assumption is acceptable for rule weighting.
  3. [Table 3] The NeuralLP row for FB15K-237 is arithmetically impossible: MRR=0.251, H@1=37.4, H@3=27.3, H@10=18.8. Under the standard definition, Hit@K is non-increasing in K, so H@1 <= H@3 <= H@10; the reported values violate this. Since this baseline is part of the comparison used to support the claim in Section 4.2 that LeSR outperforms all baselines on FB15K-237, the row must be corrected or the baseline rerun. This error also casts doubt on other baseline numbers in the same table.
  4. [Section 4.2] The claim that 'LeSR outperforms all selected baseline models across all evaluation metrics' on CN100 is contradicted by Table 6, where RotatE has a better (lower) MR of 4033.7 than LeSR's 5866.5. Similarly, on WN18RR (Table 2), LeSR GPT-3.5 has MRR=0.497, which is below AnyBURL (0.498), NeuralLP (0.509), and DRUM (0.515), so it does not rank 'among the top three across all metrics.' Such overclaims should be replaced with precise per-metric rankings.
minor comments (5)
  1. [Section 4.1] The implementation paragraph states a StepLR scheduler with 'a step size of 100 and a gamma value of 100.' A gamma of 100 would multiply the learning rate by 100 at each scheduled step, which is almost certainly a typo; please correct it (likely gamma=0.1) for reproducibility.
  2. [Throughout] The dataset name is written inconsistently as 'UMLs' and 'CN100K'/'CN100'; these should be standardized (e.g., 'UMLS' and 'CN100').
  3. [Section 3.4] The paper defines both L0 and L1 but does not state which loss is actually optimized in the experiments; please clarify the relationship between them and report the final training objective.
  4. [Appendix C] The prompt template includes example rules with NOT and OR even though Section 3.3 says the work focuses exclusively on conjunctions; this may confuse reproduction of the rule filtering step.
  5. [Section 3.3] The mapping of LLM-generated relation names to KB relations via sentence-transformer similarity is not fully specified; please report the exact mapping procedure, including any threshold or disambiguation rule, so the LLM Proposer is reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the method is a standard supervised pipeline whose rule weights and mixing coefficient are fit on training triples and evaluated on held-out triples.

full rationale

The paper's derivation chain is self-contained rather than circular. LeSR samples subgraphs from the training KB, uses an LLM to propose candidate rules, grounds those rules against the training KB to compute C_i and A_i matrices, and then learns rule significance weights w_i, the embedding weight w_emb, and the mixing factor alpha by optimizing the loss in Section 3.4. This is supervised fitting to the training KB, with evaluation on held-out test triples under standard KBC metrics; no predicted quantity is defined in terms of a fitted constant or a target result. The WD15K rule interpretability annotations used for the RCS/RQI analysis come from the external source Lv et al. (2021) and are not used to train or fit the model. There is no load-bearing self-citation: the paper's authors do not rely on their own prior work to justify the central premise, and the citation to 'He et al. 2024' in the related work section is not used as evidence for the framework and is not by the same authors. The concerns raised by the reader and skeptic, namely the potentially non-positive argument of the log in L1, the closed-world penalty on absent triples, and the arithmetically implausible NeuralLP baseline row, are internal-consistency or benchmark-validity issues rather than instances of a prediction reducing by construction to its inputs. They therefore do not count as circularity under the specified criteria.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central method rests on standard supervised learning, with rule weights fit to the training KB and evaluated on held-out triples, but it also makes an unstated closed-world assumption in training and depends on several manually chosen sampling and LLM-mapping choices that are not ablated. No new physical or mathematical entities are introduced. Confidence in the ledger is MODERATE because the paper does not release code or rule artifacts.

free parameters (4)
  • Rule significance weights w_i = Learned per rule; not reported except examples in Appendix B
    Each LLM-proposed rule receives a learnable weight optimized on the training KB with the L1 objective; these weights determine which rules dominate inference.
  • Mixture coefficient alpha and embedding weight w_emb = Learned; Appendix B shows alpha=0.980, 0.878, 0.707 for three relations
    The final score blends rule scores and RotatE with learned weights; these are fit to training data.
  • Subgraph sampling hyperparameters = Max hops = 3, max neighbors = 3, subgraphs per relation = 30
    Chosen by hand in Section 4.1 to limit LLM context; they control which rules can be proposed and no sensitivity analysis is given.
  • Sentence-transformer relation mapping = Highest-similarity mapping; no threshold reported
    In Section 3.3, generated relation names are replaced by the KB relation with highest embedding similarity; this filtering choice determines which proposed rules survive.
assumptions (5)
  • domain assumption The training KB can be treated as ground truth for rule scoring, so an absent triple is evidence against a rule.
    Section 3.4 defines s_i(h,t) = -C_i(h,t) when the rule body fires but the head triple is absent, penalizing rules that produce missing triplets. This is a closed-world assumption and conflicts with the open-world discussion in Section 2.
  • standard math Rule grounding can be computed exactly as matrix products and transposes over relation matrices.
    Appendix D gives formulas such as C = M_ri M_rj and A = C elementwise M_rk; this assumes all paths in the KB are counted correctly and that the matrices fit in memory or are sparse, which is not stated for datasets with 78K entities.
  • domain assumption The LLM's proposed rules, after filtering, cover the useful rule space for the target relation.
    The system cannot discover rules the LLM does not propose, and rules outside the 0th/1st/2nd-order taxonomy in Appendix D may be discarded, so recall of good rules depends on the LLM and the taxonomy.
  • domain assumption Sentence-transformer similarity maps LLM relation names to the correct KB relations.
    Section 3.3's second filter replaces each relation name in a proposed rule with the highest-similarity KB relation; an incorrect mapping corrupts the rule before it reaches the reasoner.
  • domain assumption Rule scores and RotatE scores are on a comparable scale and can be linearly blended.
    The final loss L1 in Section 3.4 sums alpha times rule scores with (1-alpha) times embedding scores, but no normalization of the two score types is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Model-Enhanced Symbolic Reasoning for Knowledge Base Completion." pith.science (2026). https://pith.science/paper/RTY4KAOH

@misc{pith2026250101246,
  author       = {Pith},
  title        = {Pith review of: Large Language Model-Enhanced Symbolic Reasoning for Knowledge Base Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTY4KAOH}},
  note         = {Machine review of arXiv:2501.01246}
}
read the original abstract

Integrating large language models (LLMs) with rule-based reasoning offers a powerful solution for improving the flexibility and reliability of Knowledge Base Completion (KBC). Traditional rule-based KBC methods offer verifiable reasoning yet lack flexibility, while LLMs provide strong semantic understanding yet suffer from hallucinations. With the aim of combining LLMs' understanding capability with the logical and rigor of rule-based approaches, we propose a novel framework consisting of a Subgraph Extractor, an LLM Proposer, and a Rule Reasoner. The Subgraph Extractor first samples subgraphs from the KB. Then, the LLM uses these subgraphs to propose diverse and meaningful rules that are helpful for inferring missing facts. To effectively avoid hallucination in LLMs' generations, these proposed rules are further refined by a Rule Reasoner to pinpoint the most significant rules in the KB for Knowledge Base Completion. Our approach offers several key benefits: the utilization of LLMs to enhance the richness and diversity of the proposed rules and the integration with rule-based reasoning to improve reliability. Our method also demonstrates strong performance across diverse KB datasets, highlighting the robustness and generalizability of the proposed framework.

Figures

Figures reproduced from arXiv: 2501.01246 by the authors.

Figure 1
Figure 1. LeSR: LLM-enhanced Symbolic Reasoning for KB Completion, aiming for flexibility, semantic understanding and generalizability. this challenge, the task of Knowledge Base Com￾pletion (KBC) has attracted considerable attention in the research community, aiming to automatically infer missing entities within KBs (Socher et al., 2013). Existing KBC methods are mostly embedding￾based and logic rule-based. Earlier research … view at source ↗
Figure 2
Figure 2. An overview of LeSR: LLM-enhanced Symbolic Reasoning. The Subgraph Extractor samples relevant subgraphs from the KB, then the LLM uses these subgraphs to propose logical rules, which will be further refined by the Rule Reasoner, learning the significance of the proposed rules and performing KB completion. passing to aggregate information from neighbour￾ing nodes and edges (Schlichtkrull et al., 2018). However, GNNs … view at source ↗
Figure 3
Figure 3. Numbers of proposed and learnable rules using GPT-3.5 and GPT-4. The y-axis is in log 10. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Numbers of learned logical rules of different [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 18 canonical work pages

  1. [1]

    Ivana Balazevic, Carl Allen, and Timothy Hospedales. 2019. https://doi.org/10.18653/v1/D19-1522 T uck ER : Tensor factorization for knowledge graph completion . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5185--5194...

    doi:10.18653/v1/d19-1522 2019

  2. [2]

    Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. https://proceedings.neurips.cc/paper_files/paper/2013/file/1cecc7a77928ca8133fa24680a88d2f9-Paper.pdf Translating embeddings for modeling multi-relational data . In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc

    2013

  3. [3]

    Antoine Bosselut, Hannah Rashkin, Maarten Sap, Chaitanya Malaviya, Asli Celikyilmaz, and Yejin Choi. 2019. https://doi.org/10.18653/v1/P19-1470 COMET : Commonsense transformers for automatic knowledge graph construction . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4762--4779, Florence, Italy. Associat...

    doi:10.18653/v1/p19-1470 2019

  4. [4]

    Rajarshi Das, Shehzaad Dhuliawala, Manzil Zaheer, Luke Vilnis, Ishan Durugkar, Akshay Krishnamurthy, Alex Smola, and Andrew McCallum. 2018. https://arxiv.org/abs/1711.05851 Go for a walk and arrive at the answer: Reasoning over paths in knowledge bases using reinforcement learning . Preprint, arXiv:1711.05851

    arXiv 2018

  5. [5]

    Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. https://doi.org/10.1609/aaai.v32i1.11573 Convolutional 2d knowledge graph embeddings . Proceedings of the AAAI Conference on Artificial Intelligence, 32(1)

    doi:10.1609/aaai.v32i1.11573 2018

  6. [6]

    Luis Antonio Gal\' a rraga, Christina Teflioudi, Katja Hose, and Fabian Suchanek. 2013. https://doi.org/10.1145/2488388.2488425 Amie: association rule mining under incomplete evidence in ontological knowledge bases . In Proceedings of the 22nd International Conference on World Wide Web, WWW '13, page 413–422, New York, NY, USA. Association for Computing Machinery

    arXiv 2013

  7. [7]

    Michael Galkin, Zhaocheng Zhu, Hongyu Ren, and Jian Tang. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/6246e04dcf42baf7c71e3a65d3d93b55-Paper-Conference.pdf Inductive logical query answering in knowledge graphs . In Advances in Neural Information Processing Systems, volume 35, pages 15230--15243. Curran Associates, Inc

    work page 2022

  8. [8]

    Zhongmou He, Jing Zhu, Shengyi Qian, Joyce Chai, and Danai Koutra. 2024. https://arxiv.org/abs/2406.04640 Linkgpt: Teaching large language models to predict missing links . Preprint, arXiv:2406.04640

    arXiv 2024

Show all 48 references
  1. [9]

    Hwang, Chandra Bhagavatula, Ronan Le Bras, Jeff Da, Keisuke Sakaguchi, Antoine Bosselut, and Yejin Choi

    Jena D. Hwang, Chandra Bhagavatula, Ronan Le Bras, Jeff Da, Keisuke Sakaguchi, Antoine Bosselut, and Yejin Choi. 2021. https://doi.org/10.1609/aaai.v35i7.16792 (comet-) atomic 2020: On symbolic and neural commonsense knowledge graphs . Proceedings of the AAAI Conference on Art...

    2021 doi

  2. [10]

    Tushar Khot, Sriraam Natarajan, Kristian Kersting, and Jude Shavlik. 2011. https://doi.org/10.1109/ICDM.2011.87 Learning markov logic networks via functional gradient boosting . In 2011 IEEE 11th International Conference on Data Mining, pages 320--329

    2011 doi

  3. [11]

    Stanley Kok and Pedro Domingos. 2007. https://doi.org/10.1145/1273496.1273551 Statistical predicate invention . In Proceedings of the 24th International Conference on Machine Learning, ICML '07, page 433–440, New York, NY, USA. Association for Computing Machinery

    2007

  4. [12]

    Xiang Li, Aynaz Taheri, Lifu Tu, and Kevin Gimpel. 2016. https://doi.org/10.18653/v1/P16-1137 Commonsense knowledge base completion . In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1445--1455, Berlin, G...

    2016 doi

  5. [13]

    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

  6. [14]

    Xi Victoria Lin, Richard Socher, and Caiming Xiong. 2018. https://doi.org/10.18653/v1/D18-1362 Multi-hop knowledge graph reasoning with reward shaping . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3243--3253, Brussels, Belgi...

    2018 doi

  7. [15]

    Xin Lv, Yixin Cao, Lei Hou, Juanzi Li, Zhiyuan Liu, Yichi Zhang, and Zelin Dai. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.700 Is multi-hop reasoning really explainable? towards benchmarking reasoning interpretability . In Proceedings of the 2021 Conference on Empirical...

    2021 doi

  8. [16]

    Christian Meilicke, Melisachew Wudage Chekol, Daniel Ruffinelli, and Heiner Stuckenschmidt. 2019. https://doi.org/10.24963/ijcai.2019/435 Anytime bottom-up rule learning for knowledge graph completion . In Proceedings of the Twenty-Eighth International Joint Conference on Arti...

    2019 doi

  9. [17]

    Pasquale Minervini, Sebastian Riedel, Pontus Stenetorp, Edward Grefenstette, and Tim Rocktäschel. 2020. https://arxiv.org/abs/2007.06477 Learning reasoning strategies in end-to-end differentiable proving . Preprint, arXiv:2007.06477

    2020 arXiv

  10. [18]

    Sriraam Natarajan, Tushar Khot, Kristian Kersting, Bernd Gutmann, and Jude W. Shavlik. 2010. https://api.semanticscholar.org/CorpusID:15889400 Boosting relational dependency networks . In International Conference on Inductive Logic Programming

    2010

  11. [19]

    Maximilian Nickel, Kevin Murphy, Volker Tresp, and Evgeniy Gabrilovich. 2016. https://doi.org/10.1109/JPROC.2015.2483592 A review of relational machine learning for knowledge graphs . Proceedings of the IEEE, 104(1):11--33

    2016

  12. [20]

    Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. 2011. A three-way model for collective learning on multi-relational data. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), ICML '11, pages 809--816, New York, NY, USA. ACM

    2011

  13. [21]

    Heiko Paulheim. 2017. https://doi.org/10.3233/SW-160218 Knowledge graph refinement: A survey of approaches and evaluation methods . Semantic web, 8(3):489--508

    2017 doi

  14. [22]

    Meng Qu, Junkun Chen, Louis-Pascal Xhonneux, Yoshua Bengio, and Jian Tang. 2021. https://arxiv.org/abs/2010.04029 Rnnlogic: Learning logic rules for reasoning on knowledge graphs . Preprint, arXiv:2010.04029

    2021 arXiv

  15. [23]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/D19-1410 Sentence- BERT : Sentence embeddings using S iamese BERT -networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference...

    2019 doi

  16. [24]

    Tim Rockt\" a schel and Sebastian Riedel. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/b2ab001909a8a6f04b51920306046ce5-Paper.pdf End-to-end differentiable proving . In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc

    2017

  17. [25]

    Ali Sadeghian, Mohammadreza Armandpour, Patrick Ding, and Daisy Zhe Wang. 2019. https://proceedings.neurips.cc/paper_files/paper/2019/file/0c72cb7ee1512f800abe27823a792d03-Paper.pdf Drum: End-to-end differentiable rule mining on knowledge graphs . In Advances in Neural Informa...

    2019

  18. [26]

    Smith, and Yejin Choi

    Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. 2019. https://doi.org/10.1609/aaai.v33i01.33013027 Atomic: An atlas of machine commonsense for if-then reasoning . Proceedings of the A...

    2019 doi

  19. [27]

    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. 2018. https://doi.org/10.1007/978-3-319-93417-4_38 Modeling relational data with graph convolutional networks . In The Semantic Web: 15th International Conference, ESWC 2018, ...

    2018 doi

  20. [28]

    Richard Socher, Danqi Chen, Christopher D Manning, and Andrew Ng. 2013. https://proceedings.neurips.cc/paper_files/paper/2013/file/b337e84de8752b27eda3a12363109e80-Paper.pdf Reasoning with neural tensor networks for knowledge base completion . In Advances in Neural Information...

    2013

  21. [29]

    Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. https://doi.org/10.1609/aaai.v31i1.11164 Conceptnet 5.5: An open multilingual graph of general knowledge . Proceedings of the AAAI Conference on Artificial Intelligence, 31(1)

    2017 doi

  22. [30]

    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

  23. [31]

    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

  24. [32]

    Théo Trouillon, Johannes Welbl, Sebastian Riedel, Eric Gaussier, and Guillaume Bouchard. 2016. https://proceedings.mlr.press/v48/trouillon16.html Complex embeddings for simple link prediction . In Proceedings of The 33rd International Conference on Machine Learning, volume 48 ...

    2016

  25. [33]

    Blerta Veseli, Sneha Singhania, Simon Razniewski, and Gerhard Weikum. 2023. https://doi.org/10.1007/978-3-031-33455-9_14 Evaluating language models for knowledge base completion . In The Semantic Web: 20th International Conference, ESWC 2023, Hersonissos, Crete, Greece, May 28...

    2023 doi

  26. [34]

    Quan Wang, Zhendong Mao, Bin Wang, and Li Guo. 2017. https://doi.org/10.1109/TKDE.2017.2754499 Knowledge graph embedding: A survey of approaches and applications . IEEE Transactions on Knowledge and Data Engineering, 29(12):2724--2743

    2017

  27. [35]

    Zhichun Wang and Juanzi Li. 2015. https://arxiv.org/abs/1512.07734 Rdf2rules: Learning rules from rdf knowledge bases by mining frequent predicate cycles . Preprint, arXiv:1512.07734

    2015 arXiv

  28. [36]

    Hong Wu, Zhe Wang, Kewen Wang, Pouya Ghiasnezhad Omran, and Jiangmeng Li. 2023. https://doi.org/10.4230/TGDK.1.1.7 Rule Learning over Knowledge Graphs: A Review . Transactions on Graph Data and Knowledge, 1(1):7:1--7:23

    2023 doi

  29. [37]

    Wenhan Xiong, Thien Hoang, and William Yang Wang. 2017. https://doi.org/10.18653/v1/D17-1060 D eep P ath: A reinforcement learning method for knowledge graph reasoning . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 564--573, ...

    2017 doi

  30. [38]

    Bishan Yang, Wen tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. https://arxiv.org/abs/1412.6575 Embedding entities and relations for learning and inference in knowledge bases . Preprint, arXiv:1412.6575

    2015 arXiv

  31. [39]

    Fan Yang, Zhilin Yang, and William W Cohen. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/0e55666a4ad822e0e34299df3591d979-Paper.pdf Differentiable learning of logical rules for knowledge base reasoning . In Advances in Neural Information Processing Systems,...

    2017

  32. [40]

    Yuan Yang and Le Song. 2020. https://arxiv.org/abs/1910.02481 Learn to explain efficiently via neural logic inductive learning . Preprint, arXiv:1910.02481

    2020 arXiv

  33. [41]

    Liang Yao, Jiazhen Peng, Chengsheng Mao, and Yuan Luo. 2024. https://arxiv.org/abs/2308.13916 Exploring large language models for knowledge graph completion . Preprint, arXiv:2308.13916

    2024 arXiv

  34. [42]

    Zefan Zeng, Qing Cheng, and Yuehang Si. 2023. https://doi.org/10.3390/math11214486 Logical rule-based knowledge graph reasoning: A comprehensive survey . Mathematics, 11(21)

    2023 doi

  35. [43]

    Muhan Zhang and Yixin Chen. 2018. https://proceedings.neurips.cc/paper_files/paper/2018/file/53f0d7c537d99b3824f0f99d62ea2428-Paper.pdf Link prediction based on graph neural networks . In Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc

    2018

  36. [44]

    Pan, and Huajun Chen

    Wen Zhang, Jiaoyan Chen, Juan Li, Zezhong Xu, Jeff Z. Pan, and Huajun Chen. 2022. https://arxiv.org/abs/2202.07412 Knowledge graph reasoning with logics and embeddings: Survey and perspective . Preprint, arXiv:2202.07412

    2022 arXiv

  37. [45]

    Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Wen Zhang, and Huajun Chen. 2024. https://arxiv.org/abs/2310.06671 Making large language models perform better in knowledge graph completion . Preprint, arXiv:2310.06671

    2024 arXiv

  38. [46]

    Fangfang Zhou, Jiapeng Mi, Beiwen Zhang, Jingcheng Shi, Ran Zhang, Xiaohui Chen, Ying Zhao, and Jian Zhang. 2023. https://doi.org/10.1186/s42492-023-00150-7 Reliable knowledge graph fact prediction via reinforcement learning . Visual Computing for Industry, Biomedicine, and Art, 6

    2023 doi

  39. [47]

    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...

  40. [48]

    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 10, 2026 · model on record in the stance chip above.