Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Probabilistic Circuits for Knowledge Graph Completion with Reduced Rule Sets

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A learned distribution over rule contexts lets knowledge graph completion keep 91% of peak accuracy with 70-96% fewer rules.

desk verdict A genuinely new PC-over-contexts method for rule-set reduction in KG completion, with sound formal results but an under-specified and unvalidated context-discovery step; deserves review but needs major experimental and algorithmic additions. read the letter →

arxiv 2508.06706 v1 pith:L2IXOYGO submitted 2025-08-08 cs.AI cs.LO

Jaikrishna Manojkumar Patil , Nathaniel Lee , Al Mehdi Saadat Chowdhury , YooJung Choi , Paulo Shakarian This is my paper
classification cs.AIcs.LO
keywords probabilisticcircuitsknowledgegraphcompletionrule-basedreasoningrulecontextssetreductionAnyBURLexplainabilitylogic
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 that the reason rule-based knowledge graph completion needs tens of thousands of rules is that confidence-based selection ignores which rules work together. The authors introduce rule contexts—meaningful subsets of rules that jointly entail queries—discovered from the training graph, learn a probability distribution over contexts with a probabilistic circuit, and then answer queries from small rule subsets chosen by marginal probability. They report a 70-96% reduction in rules (about a 12-fold reduction) while preserving on average 91% of the best baseline accuracy, and up to a 31-fold speedup in reaching a given accuracy. The point is not just compression: the reduced rule sets keep the reasoning chain legible, restoring the explainability that rule-based methods promise.

What carries the argument

The central object is a 'rule context': a mutually exclusive mode in which some learned first-order rules activate and others do not, encoded with activation atoms $\mu_r$ and context atoms $\nu_c$. A probabilistic circuit learns $P_\theta$ over these contexts; marginals over rule sets are obtained by summing context probabilities, with Propositions 1-4 giving lower bounds, exact probabilities, and upper bounds without any independence assumption.

What would settle it

Take one benchmark dataset, split the training triples into two halves, build contexts and learn the circuit on each half, and compare the top-K rule subsets selected by marginal probability. If the two halves select substantially different rule subsets, or if subsets selected from one half lose most of the claimed 91% accuracy when evaluated on the other half's queries, then the discovered contexts are not carrying the generalization.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a probabilistic circuit learned over rule contexts captures rule interactions that confidence scores miss, so the marginal probability of a rule set is a better guide for selecting rules than individual rule confidence. Formally, for a learned distribution $P_\theta$ over mutually exclusive contexts, the marginal of a query $q$ equals the sum of $P_\theta(c)$ over contexts that entail $q$; the paper proves that this equals $1 - P_\theta(\bigwedge_{c: \Pi_c \models q} \neg\nu_c)$, giving an exact query probability, and derives lower and upper bounds when only subsets of rules are available. Experiments on eight benchmark knowledge graphs

Load-bearing premise

The load-bearing premise is that the rule contexts produced from training triples by the abductive engine are the meaningful groups of rules and that the circuit's learned distribution over them continues to hold on unseen test queries; if the contexts are arbitrary or training-biased, the reduced rule sets will not transfer.

Editorial extensions

If this is right

  • Rule-based knowledge graph completion can be made explainable at scale: each prediction is supported by a handful of rules rather than thousands.
  • Selection by learned marginal probability over contexts outperforms confidence-based rule selection at equal rule counts, often by large multiples.
  • Because no independence assumption is used and the semantics align with Nilsson's probabilistic logic, the framework offers a general way to shrink rule sets in any rule-based reasoner.
  • The same machinery can be applied beyond knowledge graphs, for example to inductive logic programming or proof search, where rule explosion is also a bottleneck.

Reading between the lines

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

  • The 91% preservation is measured against the confidence-based baseline's full rule set; a sterner test would compare against embedding-based completers or against the same circuit using all rules at an equal inference budget.
  • Because contexts are discovered from training triples by an abductive engine, the method's transfer depends on whether those contexts are stable across training splits; one could test this by learning contexts on subsamples and checking how much the selected rule subsets change.
  • The framework could be turned into a rule-learning guide, using the learned distribution over contexts to propose which candidate rules to replace rather than only which ones to keep.
  • The circuit's structure over contexts may itself be inspected to explain why a cluster of rules is predictive, for example by identifying shared body atoms that drive high context probability.
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 paper proposes a probabilistic-circuit-guided framework for knowledge graph completion. It introduces 'rule contexts' (subsets of learned rules that work together), learns a PC over those contexts, and uses marginal probabilities to rank and select small rule sets. Three inference methods are presented: PC1 (lower-bound approximation with singleton rule sets), PC2 (exact query probability), and PC3 (lower-bound approximation with greedy-walk rule sets). The authors provide formal propositions about these marginals and a semantics connection to Nilsson probabilistic logic. Experiments on eight KG benchmarks claim a 70–96% reduction in rules, a 31-fold performance improvement over the baseline at equal small rule counts, and preservation of about 91% of peak baseline performance.

Significance. If the empirical claims hold, the paper offers a promising way to make rule-based KG completion more explainable by using a PC to select far fewer rules, without assuming rule independence, and with tractable exact/lower-bound inference. The inclusion of formal semantics and the release of code/datasets in the supplementary material are strengths. However, the central context-discovery step is not described or validated, several formal statements are incorrect as written, and the experimental protocol is too thin to support the headline quantitative claims. The contribution is therefore interesting but not yet established.

major comments (4)
  1. [Technical Preliminaries, Proposition 1] As written, Proposition 1 is false. Take R={p←, q←} and Πc={p←} with Pθ(c)>0, and suppose no context entails q←. Then Πc ⊆ R, but Πc does not entail R, so the left-hand side Pθ(R) can be 0 while the right-hand side is at least Pθ(c)>0. The proof's statement 'if Πc is contained in R then R must be a logical consequence' is not a valid inference. The intended statement may be with R ⊆ Πc, but as written this is a load-bearing formal error.
  2. [Technical Preliminaries, Proposition 4] Proposition 4 is also false as stated. Let Rj=∅. For any non-tautological query q, Rj does not entail q, and Pθ(Rj)=1 because every program entails the empty set. Then the claimed upper bound becomes Pθ(q) ≤ 0, which fails whenever Pθ(q)>0. The correct upper-bound condition should ensure disjointness, e.g., Rj |= ¬q (or a similar condition). This matters because the paper advertises Propositions 1–4 as a suite of formal guarantees, and a reader cannot verify which results are actually used by the proposed inference methods.
  3. [Rule Context and Experimental Setup] The rule-context discovery step is the unvalidated crux of the framework. The paper only says 'We use training triples to establish contexts' and 'We have used PyClause's abductive reasoning engine to determine rule-context associations,' with no definition of a context, no algorithm or configuration, no number of resulting contexts, and no consistency check. Since Eq. (1) and all PC marginals are conditioned on this association matrix, the claimed 70–96% rule reduction and performance preservation could be artifacts of this step. Please provide a precise description and a sensitivity/robustness analysis of the context-generation procedure.
  4. [Experimental Setup and Results and Discussion] The experimental protocol is too thin to support the quantitative claims. Seven of the eight datasets are run once (only Family has 5 runs), and hyperparameters such as the confidence threshold and number of EM iterations are chosen per dataset. The headline comparisons ('31-fold improvement', '91% preservation', 'minimal equivalent number of rules') are not formally defined, and the WN18 baseline has zero performance at 500 rules, which makes fold-ratio comparisons unbounded. Please report per-dataset raw values, define the aggregation precisely, and provide multiple seeds or error bars.
minor comments (5)
  1. [Formal Probabilistic Semantics, Theorem 1] The theorem statement uses R on the right-hand side ('Σ_{c s.t. Πc|=R} Pθ(c)') but the theorem is about a query q. This should be q, not R.
  2. [Proposition 2 proof] The proof writes '¬c′″' where it should write '¬νc′″'; the helper atom symbol is missing.
  3. [Appendix A, Algorithm 1] The indentation of Algorithm 1 appears incorrect: lines 8–9 (updating S and R_remaining) are inside the inner while loop, and 'return S' appears before the outer while loop terminates. Please fix the loop structure.
  4. [Appendix B, Figure 11] The caption says 'UML dataset' but the dataset name in Table 1 is UMLS. Please correct the typo.
  5. [Hyperparameter Selection] For WN18, WN18RR, and Family, the 'confidence threshold' is reported as 0%. Since AnyBURL's confidence is a ratio, a 0% threshold typically means no filtering; please clarify whether this means all mined rules were used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: formal results are definitional identities and empirical claims are held-out test measurements.

full rationale

The paper's derivation chain is self-contained. Propositions 1-4 and Theorem 1 are proven from the definition of Pθ(R)=Σ_{c:Πc|=R}Pθ(c) and the mutual-exclusivity of contexts; they do not presuppose the empirical rule-reduction claim. The central empirical claims (70-96% rule reduction, ~91% performance preservation) compare PC-selected rule subsets against a confidence-based baseline on held-out test triples. The PC marginals are learned from training triples only (via EM on contexts discovered by PyClause), while Hits@k/MRR are computed on the test set with AnyBURL's evaluation engine—so the reported performance is a separate measurement, not a restatement of the fitted marginals. Self-citations (Choi et al. 2020 for PC definitions, Shakarian et al. 2012 for Nilsson-style semantics, Aditya et al. 2023 for first-order syntax) are standard references used for background, not to justify the novelty. The underspecified PyClause context-generation step is a reproducibility/validity limitation, not a circular reduction; no equation in the paper forces the test results from the training-fit quantities.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim rests on the definition of contexts as discovered rule subsets, the mutual exclusivity and exhaustiveness of contexts, and the determinism of rule entailment within a context. The only fitted numbers are per-dataset confidence thresholds and EM iterations, which are chosen by the authors to favor the method. The invented entities (contexts, activation atoms) are motivated by the approach but lack independent evidence.

free parameters (5)
  • Confidence threshold for input rules = Per dataset: 50% (CODEX-S/Kinship/UMLS), 60% (FB15K-237), 70% (Nations), 0% (WN18, WN18RR, Family)
    Chosen per dataset to control the size of the input rule set. The paper states 'The final settings focused on showing framework effectiveness over large number of rules for each dataset', indicating selection bias.
  • Number of EM iterations for PC learning = 100 (Family/WN18/WN18RR), 50 (Kinship), 10 (others)
    Chosen per dataset without a principled criterion; affects the quality of the learned circuit and thus the rule marginals.
  • AnyBURL learning time = 10 seconds
    Fixed for all datasets, but arbitrary; the number of learned rules depends heavily on this.
  • AnyBURL minimum support threshold = >= 10
    Arbitrary cutoff that influences the rule set.
  • Greedy walk threshold delta for PC3 = Not specified in the paper
    Used in Algorithm 1 to terminate walks; no value is reported, making PC3 results irreproducible.
assumptions (4)
  • domain assumption Contexts are mutually exclusive and exhaustive: every world with non-zero probability satisfies exactly one context atom νc.
    Used in Lemma 1 (Claim 1) and Proposition 2. The paper adds rules for mutual exclusivity but does not add rules to enforce that at least one context holds; this is assumed via Pθ summing to 1.
  • domain assumption For a context c, if Πc entails q, then q is true in that context with probability 1.
    The entire framework treats rule entailment as deterministic within a context. Real rules have confidence scores, so this is a simplifying assumption.
  • domain assumption The PyClause abductive reasoning engine correctly identifies the rule-context associations from training data.
    The paper relies on this external tool to generate the binary matrix M, which is the training data for the PC. No formal guarantee or detailed algorithm is provided.
  • standard math Standard first-order logic semantics with worlds as subsets of ground atoms.
    The paper uses standard Tarskian semantics for rules and entailment in the technical preliminaries.
invented entities (2)
  • Rule contexts
    purpose: Subsets of rules that work together, used to learn a distribution and select compact rule sets.
    Contexts are constructed from training data via abduction; there is no external or falsifiable handle independent of the paper's own pipeline.
  • Activation atoms μr and context atoms νc
    purpose: Auxiliary logical constructs to formalize the association of rules with contexts in the logic program.
    These are introduced purely as internal scaffolding for the formalism; they have no observable counterpart.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic Circuits for Knowledge Graph Completion with Reduced Rule Sets." pith.science (2026). https://pith.science/paper/L2IXOYGO

@misc{pith2026250806706,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Circuits for Knowledge Graph Completion with Reduced Rule Sets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L2IXOYGO}},
  note         = {Machine review of arXiv:2508.06706}
}
abstract

Rule-based methods for knowledge graph completion provide explainable results but often require a significantly large number of rules to achieve competitive performance. This can hinder explainability due to overwhelmingly large rule sets. We discover rule contexts (meaningful subsets of rules that work together) from training data and use learned probability distribution (i.e. probabilistic circuits) over these rule contexts to more rapidly achieve performance of the full rule set. Our approach achieves a 70-96% reduction in number of rules used while outperforming baseline by up to 31$\times$ when using equivalent minimal number of rules and preserves 91% of peak baseline performance even when comparing our minimal rule sets against baseline's full rule sets. We show that our framework is grounded in well-known semantics of probabilistic logic, does not require independence assumptions, and that our tractable inference procedure provides both approximate lower bounds and exact probability of a given query. The efficacy of our method is validated by empirical studies on 8 standard benchmark datasets where we show competitive performance by using only a fraction of the rules required by AnyBURL's standard inference method, the current state-of-the-art for rule-based knowledge graph completion. This work may have further implications for general probabilistic reasoning over learned sets of rules.

Figures

Figures reproduced from arXiv: 2508.06706 by the authors.

Figure 1
Figure 1. Rule set explosion in AnyBURL across benchmark [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Hits@10 performance of PC1 and PC2 inferences against Baseline. PC variants attained competitive performance for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Mean rank reciprocal (MRR) performance of PC1 and PC2 inferences against Baseline. PC variants attained compet [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 7
Figure 7. Figure 7: Runtime in seconds of PC1(left) and PC2(right) [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 5
Figure 5. Figure 5: Runtime scalability for generating context rules [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Runtime comparison for singleton(left) vs greedy [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Greedy walks performance on Family dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 11
Figure 11. Figure 11: Greedy walks performance on UML dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 10
Figure 10. Figure 10: Greedy walks performance on Kinship dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 12
Figure 12. Figure 12: Hits@1 performance of PC1 and PC2 inferences against Baseline. PC variants attained competitive performance for [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Hits@3 performance of PC1 and PC2 inferences against Baseline. PC variants attained competitive performance for [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Aditya, D.; Mukherji, K.; Balasubramanian, S.; Chaudhary, A.; and Shakarian, P. 2023. Pyreason: Software for open world temporal logic. arXiv preprint arXiv:2302.13482

    arXiv 2023

  2. [2]

    Ahmed, K.; Chang, K.-W.; and Van den Broeck, G. 2024. Controllable generation via locally constrained resampling. arXiv preprint arXiv:2410.13111

    arXiv 2024

  3. [3]

    Ahmed, K.; Teso, S.; Chang, K.-W.; Van den Broeck, G.; and Vergari, A. 2022 a . Semantic Probabilistic Layers for Neuro-Symbolic Learning. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A., eds., Advances in Neural Information Processing Systems, volume 35, 29944--29959. Curran Associates, Inc

    work page 2022

  4. [4]

    Ahmed, K.; Wang, E.; Chang, K.-W.; and Van den Broeck, G. 2022 b . Neuro-symbolic entropy regularization. In Uncertainty in Artificial Intelligence, 43--53. PMLR

    work page 2022

  5. [5]

    d.; Serafini, L.; and Spranger, M

    Badreddine, S.; Garcez, A. d.; Serafini, L.; and Spranger, M. 2022. Logic tensor networks. Artificial Intelligence, 303: 103649

    work page 2022

  6. [6]

    M.; and Stuckenschmidt, H

    Betz, P.; Galarraga, L.; Ott, S.; Meilicke, C.; Suchanek, F. M.; and Stuckenschmidt, H. 2024. PyClause-Simple and Efficient Rule Handling for Knowledge Graphs. In IJCAI, demo track. Ijcai.org

    work page 2024

  7. [7]

    Bordes, A.; Usunier, N.; Garcia-Duran, A.; Weston, J.; and Yakhnenko, O. 2013 a . Translating Embeddings for Modeling Multi-relational Data. In Burges, C.; Bottou, L.; Welling, M.; Ghahramani, Z.; and Weinberger, K., eds., Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc

    work page 2013

  8. [8]

    Bordes, A.; Usunier, N.; Garcia-Duran, A.; Weston, J.; and Yakhnenko, O. 2013 b . Translating embeddings for modeling multi-relational data. Advances in neural information processing systems, 26

    work page 2013

Show all 46 references
  1. [9]

    Cheng, K.; Ahmed, N.; and Sun, Y. 2023 a . Neural Compositional Rule Learning for Knowledge Graph Reasoning. In The Eleventh International Conference on Learning Representations

    2023

  2. [10]

    K.; and Sun, Y

    Cheng, K.; Ahmed, N. K.; and Sun, Y. 2023 b . Neural compositional rule learning for knowledge graph reasoning. arXiv preprint arXiv:2303.03581

    2023 arXiv

  3. [11]

    Choi, Y.; Vergari, A.; and Van den Broeck, G. 2020. Probabilistic circuits: A unifying framework for tractable probabilistic models. UCLA. URL: http://starai. cs. ucla. edu/papers/ProbCirc20. pdf, 6

    2020

  4. [12]

    C \^o rte-Real, J.; Dries, A.; Dutra, I.; and Rocha, R. 2018. Improving candidate quality of probabilistic logic models. In Technical Communications of the 34th International Conference on Logic Programming (ICLP 2018), 6--1. Schloss Dagstuhl--Leibniz-Zentrum f \"u r Informatik

    2018

  5. [13]

    Dang, M.; Khosravi, P.; Liang, Y.; Vergari, A.; and Van den Broeck, G. 2021. Juice: A julia package for logic and probabilistic circuits. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, 16020--16023

    2021

  6. [14]

    De Raedt, L.; Dries, A.; Thon, I.; Van den Broeck, G.; and Verbeke, M. 2015. Inducing probabilistic relational rules from probabilistic examples. In Proceedings of 24th international joint conference on artificial intelligence (IJCAI), volume 2015, 1835--1842. IJCAI-INT JOINT ...

    2015

  7. [15]

    Dettmers, T.; Minervini, P.; Stenetorp, P.; and Riedel, S. 2018 a . Convolutional 2d knowledge graph embeddings. In Proceedings of the AAAI conference on artificial intelligence, volume 32

    2018

  8. [16]

    Dettmers, T.; Minervini, P.; Stenetorp, P.; and Riedel, S. 2018 b . Convolutional 2D knowledge graph embeddings. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eight...

    2018

  9. [17]

    Feldstein, J.; Phillips, D.; and Tsamoura, E. 2024. Efficiently learning probabilistic logical models by cheaply ranking mined rules. arXiv preprint arXiv:2409.16238

    2024

  10. [18]

    Gal \'a rraga, L.; Teflioudi, C.; Hose, K.; and Suchanek, F. M. 2015. Fast rule mining in ontological knowledge bases with AMIE + . The VLDB Journal, 24(6): 707--730

    2015

  11. [19]

    A.; Teflioudi, C.; Hose, K.; and Suchanek, F

    Gal \'a rraga, L. A.; Teflioudi, C.; Hose, K.; and Suchanek, F. 2013. AMIE: association rule mining under incomplete evidence in ontological knowledge bases. In Proceedings of the 22nd international conference on World Wide Web, 413--422

    2013

  12. [20]

    Giunchiglia, E.; and Lukasiewicz, T. 2020. Coherent hierarchical multi-label classification networks. Advances in neural information processing systems, 33: 9662--9673

    2020

  13. [21]

    Hayes, P. J. 1977. In defence of logic. In Proc. IJCAI-77, 559--565

    1977

  14. [22]

    Jain, A.; Friedman, T.; Kuzelka, O.; Van den Broeck, G.; and De Raedt, L. 2019. Scalable rule learning in probabilistic knowledge bases. Automated Knowledge Base Construction

    2019

  15. [23]

    Jiang, D.; Fonseca, M.; and Cohen, S. B. 2024. Leanreasoner: Boosting complex logical reasoning with lean. arXiv preprint arXiv:2403.13312

    2024 arXiv

  16. [24]

    V.; Nau, D.; Sliva, A.; Simari, G

    Khuller, S.; Martinez, M. V.; Nau, D.; Sliva, A.; Simari, G. I.; and Subrahmanian, V. S. 2007. Computing most probable worlds of action probabilistic logic programs: scalable estimation for 10\^ 30,000 worlds. Annals of Mathematics and Artificial Intelligence, 51(2-4): 295--331

    2007

  17. [25]

    Kok, S.; and Domingos, P. 2007. Statistical predicate invention. In Proceedings of the 24th International Conference on Machine Learning, ICML '07, 433–440. New York, NY, USA: Association for Computing Machinery. ISBN 9781595937933

    2007

  18. [26]

    Liu, A.; and Van den Broeck, G. 2021. Tractable regularization of probabilistic circuits. Advances in Neural Information Processing Systems, 34: 3558--3570

    2021

  19. [27]

    Liu, C.; Wei, Z.; and Zhou, L. 2025. Contrastive Predictive Embedding for learning and inference in knowledge graph. Knowledge-Based Systems, 307: 112730

    2025

  20. [28]

    Loconte, L.; Di Mauro, N.; Peharz, R.; and Vergari, A. 2023. How to turn your knowledge graph embeddings into generative models. Advances in Neural Information Processing Systems, 36: 77713--77744

    2023

  21. [29]

    Manhaeve, R.; Dumancic, S.; Kimmig, A.; Demeester, T.; and De Raedt, L. 2018. DeepProbLog: Neural Probabilistic Logic Programming. In Bengio, S.; Wallach, H.; Larochelle, H.; Grauman, K.; Cesa-Bianchi, N.; and Garnett, R., eds., Advances in Neural Information Processing System...

    2018

  22. [30]

    W.; Fink, M.; and Stuckenschmidt, H

    Meilicke, C.; Chekol, M. W.; Fink, M.; and Stuckenschmidt, H. 2020. Reinforced anytime bottom up rule learning for knowledge graph completion. arXiv preprint arXiv:2004.04412

    2020 arXiv

  23. [31]

    Q.; Nguyen, T

    Nguyen, D. Q.; Nguyen, T. D.; Nguyen, D. Q.; and Phung, D. 2017. A novel embedding model for knowledge base completion based on convolutional neural network. arXiv preprint arXiv:1712.02121

    2017 arXiv

  24. [32]

    Nilsson, N. J. 1986. Probabilistic Logic. Artificial Intelligence, 28(1): 71--87

    1986

  25. [33]

    G.; Wang, K.; and Wang, Z

    Omran, P. G.; Wang, K.; and Wang, Z. 2018. Scalable rule learning via learning representation. In IJCAI-18. International Joint Conferences on Artificial Intelligence

    2018

  26. [34]

    Ott, S.; Meilicke, C.; and Samwald, M. 2021. SAFRAN : An interpretable, rule-based link prediction method outperforming embedding models. In 3rd Conference on Automated Knowledge Base Construction

    2021

  27. [35]

    Reiter, R. 1980. A logic for default reasoning. Artificial intelligence, 13(1-2): 81--132

    1980

  28. [36]

    Safavi, T.; and Koutra, D. 2020. C o DE x: A C omprehensive K nowledge G raph C ompletion B enchmark. In Webber, B.; Cohn, T.; He, Y.; and Liu, Y., eds., Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 8328--8350. Online: Associa...

    2020

  29. [37]

    I.; and Subrahmanian, V

    Shakarian, P.; Simari, G. I.; and Subrahmanian, V. S. 2012. Annotated Probabilistic Temporal Logic: Approximate Fixpoint Implementation. ACM Transactions on Computational Logic, 13(2): 13:1--13:33

    2012

  30. [38]

    S.; and Kersting, K

    Skryagin, A.; Stammer, W.; Ochs, D.; Dhami, D. S.; and Kersting, K. 2021. SLASH: embracing probabilistic circuits into neural answer set programming. arXiv preprint arXiv:2110.03395

    2021 arXiv

  31. [39]

    Sun, Z.; Deng, Z.-H.; Nie, J.-Y.; and Tang, J. 2019. Rotate: Knowledge graph embedding by relational rotation in complex space. arXiv preprint arXiv:1902.10197

    2019 arXiv

  32. [40]

    Toutanova, K.; Chen, D.; Pantel, P.; Poon, H.; Choudhury, P.; and Gamon, M. 2015. Representing text for joint embedding of text and knowledge bases. In Proceedings of the 2015 conference on empirical methods in natural language processing, 1499--1509

    2015

  33. [41]

    Trouillon, T.; Welbl, J.; Riedel, S.; Gaussier, \'E .; and Bouchard, G. 2016. Complex embeddings for simple link prediction. In International conference on machine learning, 2071--2080. PMLR

    2016

  34. [42]

    u nchmeyer, J.; Leser, U.; and Rockt \

    Weber, L.; Minervini, P.; M \"u nchmeyer, J.; Leser, U.; and Rockt \"a schel, T. 2019. Nlprolog: Reasoning with weak unification for question answering in natural language. arXiv preprint arXiv:1906.06187

    2019 arXiv

  35. [43]

    Xiong, B.; Nayyeri, M.; Luo, L.; Wang, Z.; Pan, S.; and Staab, S. 2024. NestE: modeling nested relational structures for knowledge graph reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 9205--9213

    2024

  36. [44]

    Xu, J.; Zhang, Z.; Friedman, T.; Liang, Y.; and Van den Broeck, G. 2018. A semantic loss function for deep learning with symbolic knowledge. In International conference on machine learning, 5502--5511. PMLR

    2018

  37. [45]

    Yang, F.; Yang, Z.; and Cohen, W. W. 2017. Differentiable learning of logical rules for knowledge base reasoning. Advances in neural information processing systems, 30

    2017

  38. [46]

    Zhang, R.; Mao, Y.; and Zhao, W. 2020. Knowledge graphs completion via probabilistic reasoning. Information Sciences, 521: 144--159

    2020

Pith tools

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