Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Knowledge Graph Embeddings with Representing Relations as Annular Sectors

T0 review · 4 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 SectorE models knowledge-graph relations as annular sectors in polar coordinates, where bumped entity points inside the sectors score as true triples, and reports competitive or best link-prediction results on three benchmarks.

desk verdict SectorE is a plausible new relation geometry with a fixable flaw in its continuity constant and only marginal empirical gains. read the letter →

arxiv 2506.11099 v1 pith:GLLX44AF submitted 2025-06-06 cs.LG cs.AIcs.CL

Huiling Zhu , Yingqi Zeng This is my paper
classification cs.LGcs.AIcs.CL
keywords knowledgegraphcompletionlinkpredictionregion-basedembeddingspolarcoordinatesannularsectorssemantichierarchyrelationpatterns
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

SectorE is a knowledge graph embedding that represents each relation as an annular sector — a ring segment with inner and outer radius and an angular span — in polar coordinates, and represents each entity as a point whose position is shifted by a learned bump from the partner entity in the triple. The paper's aim is to capture two things at once: the semantic hierarchy of entities through the radial (modulus) part, and logical relation patterns through the angular (phase) part. The authors argue that annular sectors can express symmetry, antisymmetry, inversion, subsumption, intersection, and mutual exclusion simply by relating the sectors geometrically, and they report that SectorE achieves the best MRR and Hits@1 among compared baselines on YAGO3-10, with competitive results on FB15k-237 and WN18RR. A reader should care because a single geometric representation that encodes both entity hierarchy and relation logic would make link prediction more faithful to the structure of real knowledge graphs.

What carries the argument

The annular sector is the central object: a region in polar coordinates bounded by a radial interval $[l(r),u(r)]$ and an angular interval $[\psi(r),\phi(r)]$, with central modulus $c(r)$, annular depth $w(r)$, and central angle $\delta(r)$. Each relation is represented by two such sectors, one for the head position and one for the tail position. Entities are pairs of a base vector and a transformation bump; the bump operation multiplies the moduli and adds the phases modulo $2\pi$, so an entity's position depends on which partner it appears with. The scoring function combines a modulus distance and a sine-based phase distance from the bumped point to the sector center, using depth and angle factors to push points toward the region. Containment of bumped points in the sectors defines truth, sector inclusion expresses subsumption and intersection, and sector disjointness expresses antisymmetry and mutual exclusion.

What would settle it

Train SectorE on a synthetic knowledge graph containing only the subsumption pattern $r_1(x,y) \to r_2(x,y)$, then check whether the learned head sectors satisfy $R^h_1 \subseteq R^h_2$ and the tail sectors satisfy $R^t_1 \subseteq R^t_2$, or equivalently whether every true $r_1$ triple outranks every $r_2$-only triple. If the inclusion fails, the claimed pattern capture is a heuristic rather than a proven property; alternatively, evaluate the phase-distance function across a sector boundary for a relation whose central angle is not the special value, and check whether the distance jump matches the paper's continuity constant.

Watch

Extended reading notes

Core claim

The central claim is that relations are better modeled as two-dimensional polar regions — annular sectors — than as points, hyperplanes, or boxes, because the sector's radius channel and angle channel can carry different kinds of semantic information at once. In SectorE, each relation $r$ owns a head sector and a tail sector; a triple $r(h,t)$ is considered true when the head point, after being bumped by the tail's transformation, lies in the head sector and the tail point, after being bumped by the head's transformation, lies in the tail sector. The scoring function measures the negative distance from the bumped points to the sector centers, split into a modulus distance and a phase distance, so training pulls correct triples inside the sectors and pushes incorrect triples out. On top of matching or exceeding existing translational and semantic matching baselines, the learned sector areas show interpretable behavior: relations with semantically diverse tails get larger sectors, and the ratio of head to tail area tracks relation cardinality. That is the discovery: a polar annular-sector geometry can carry hierarchy and relation patterns in one embedding.

Load-bearing premise

The pattern-learning guarantees assume exact containment of bumped entity points in the annular sectors, while the actual scoring function is a soft distance that only approximates containment, and the phase-distance continuity constant is exact for only one central angle.

Editorial extensions

If this is right

  • Because symmetry, antisymmetry, inversion, subsumption, intersection, and mutual exclusion all reduce to simple sector relations, a trained SectorE model can in principle exhibit several logically distinct patterns without architectural changes.
  • The modulus channel gives the model a built-in notion of entity hierarchy, so entities at different semantic levels (like 'animal' versus 'dog') can occupy different radial bands rather than being treated as interchangeable points.
  • The reported YAGO3-10 results, where SectorE leads in MRR and Hits@1, indicate that region-based models are especially effective on large graphs with many entities per relation.
  • The observed correlation between head-to-tail sector area ratio and relation cardinality means the learned geometry can serve as an interpretability tool, exposing whether a relation is 1-N, N-1, or N-N after training.
  • Removing the transformational bump degrades performance across all metrics, confirming that the partner-dependent displacement is doing essential work in placing entities inside the correct sectors.

Reading between the lines

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

  • The pattern proofs treat sector containment as exact truth, but the actual scorer is a soft distance, so I would not expect trained sectors to literally satisfy the subset and disjointness conditions; testing those conditions on a trained model would show how much of the claimed logical expressivity survives optimization.
  • The phase-distance continuity constant is exact only for one central angle, which means gradients are likely discontinuous at sector boundaries for most relations; a corrected phase-distance formula could make training smoother and possibly improve the reported numbers.
  • The sector-area and cardinality correlation suggests a cheap post-hoc test: train on a graph with known 1-N and N-1 relations and see whether the area ratios separate them cleanly, which would turn the observation into a usable diagnostic.
  • The paper does not analyze composition patterns, a known weakness of BoxE; whether annular sectors can also express composition would delimit the representational power of this geometry.
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 SectorE, a knowledge graph embedding model in which entities are embedded as points in polar coordinates and relations are represented as annular sectors (a head sector and a tail sector). The position of an entity in a triple is obtained by combining a basic entity vector with a 'bump' vector from the other entity, using elementwise modulus multiplication and phase addition. The scoring function measures the distance from the bumped entity points to the corresponding sector, split into a modulus part and a phase part, and is trained with self-adversarial negative sampling. The authors claim that the model can express symmetry, antisymmetry, inversion, subsumption, intersection, and mutual exclusion through set-theoretic containment and disjointness of sectors, and they report link prediction results on FB15k-237, WN18RR, and YAGO3-10, with competitive or best MRR on YAGO3-10. Ablation studies show that the bump, modulus, and phase components each contribute to performance.

Significance. If the theoretical claims held, SectorE would be a useful addition to region-based KGE models, combining hierarchy awareness from polar coordinates with the expressibility of annular regions. The paper makes a genuine design contribution: the annular-sector geometry and the entity-bump interaction are clearly described and are visually intuitive. The empirical results are competitive on standard benchmarks, and the ablation study is helpful. However, the central theoretical claim—that the model provably captures six relation patterns—currently rests on a hard set-membership semantics that the soft scoring function does not implement, and the phase-distance function of Eq. (4) is not continuous for generic δ. These issues are load-bearing because the pattern proofs are the main evidence for the model's expressive power. The paper is of interest to the KGE community, but the theoretical and experimental presentation needs substantial revision before the claims can be accepted.

major comments (4)
  1. [Section III-B-3, Eq. (4)] The phase distance function is not continuous at the sector boundary for generic δ. With k = 0.5(δ − 1/δ), the inside-branch value at the boundary p = θ ± δ/2 is sin(δ/4)/δ, while the outside-branch value is sin(δ/4)·δ − 0.5(δ − 1/δ). These are equal only when δ = 1 or sin(δ/4) = 0.5. For example, with δ = π/2, the inside value is about 0.244 and the outside value is about 0.134, so a point just outside the sector receives a higher score than a point on the boundary. This discontinuity is not merely cosmetic: it means the soft scoring function does not realize the 'inside the sector gets a better score' semantics that the pattern proofs in Section III-D assume. The authors should either correct the constant to sin(δ/4)(δ − 1/δ), or provide a different continuity/score-separation argument.
  2. [Section III-D] The pattern-capturing proofs (symmetry, antisymmetry, inversion, subsumption, intersection, mutual exclusion) are stated purely in terms of set membership: a triple holds iff the bumped entity points lie inside the corresponding sectors. The actual scoring function is a soft distance plus margin and is trained with negative sampling; it never explicitly tests membership. Even if Eq. (4) were corrected, the proofs would only show that the geometry can represent the patterns at the level of region containment, not that the trained scoring function implements this membership relation. The paper needs either a formal statement of how scores are converted to truth values (e.g., a margin condition that separates inside from outside points) or a clearly scoped heuristic claim about pattern capture. As written, the proofs do not establish what the model actually learns.
  3. [Section IV-A-3 and Section III-B-2] The hyperparameter β appears in the definition of the central angle δ(r) = ϕ(r) − ψ(r) + βπ, but the grid search ranges in Section IV-A-3 do not list β. Since δ directly controls the phase distance and the continuity constant in Eq. (4), the omission makes the experimental setup incomplete and prevents reproduction. The authors should specify the range and chosen value of β, and also clarify whether β is a scalar shared by all relations or per-relation. Similarly, the text says λ1 and λ2 are learned, but the training details do not state how their initialization or regularization is handled.
  4. [Section III-C, Eq. (7)] The negative-sampling probability in Eq. (7) is written as p(h'_j, r, t'_j | ...) = α · score(...) / Σ_i exp(α · score(...)). The numerator is missing the exponential function; as written, p is not a probability (it can be negative when the score is negative, and it does not sum to one unless α·score happens to be an unnormalized log-probability). This appears to be a typo for exp(α · score(...)), but it should be corrected because the loss function in Eq. (6) depends on this distribution.
minor comments (5)
  1. [Section III-D-5] The intersection pattern condition contains a typo: the second conjunct is written as ((r_h^1 ∩ r_h^2) ⊆ r_h^3), but it should be a condition on the tail sectors, presumably ((r_t^1 ∩ r_t^2) ⊆ r_t^3).
  2. [Section IV-B, Table II] The TransE row in Table II has missing values for FB15k-237 and YAGO3-10, and the table uses inconsistent dashes; the authors should either fill in all baseline numbers from the original papers or mark them clearly as unavailable.
  3. [Section IV-D, Table IV] The heading 'anunlar sectors' is a typo for 'annular sectors'.
  4. [General] The paper contains several formatting artifacts: the index terms 'component, formatting, style, styling, insert' are leftover from a template; 'Y AGO3-10' appears with an extra space throughout; and Eq. (5) has a grammatical typo ('the total scoring function of is').
  5. [Section IV-A-4] The baseline list omits recently proposed region-based models that are discussed in the related work, such as ExpressivE and Octagon embeddings; including them (or explaining why they are excluded) would strengthen the empirical comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SectorE's pattern capabilities are constructive set-theoretic proofs over its defined region semantics, and its empirical results are benchmark evaluations, not predictions fitted to themselves.

full rationale

The paper's central derivations are (i) the construction of annular-sector embeddings with a soft distance scoring function (Eqs. 3-5), (ii) Section III-D proofs that particular sector configurations satisfy symmetry, antisymmetry, inversion, subsumption, intersection, and mutual exclusion, and (iii) supervised benchmark evaluation. None of these reduce to its inputs. The Section III-D arguments are direct set-theoretic constructions under the paper's explicit membership semantics ('ri(x, y) holds iff eri(x,y)_x ∈ rh_i and eri(x,y)_y ∈ rt_i'), with existence shown by setting sectors equal, swapped, disjoint, or included; this is the standard way to demonstrate expressiveness, not a fitted quantity renamed as a prediction. The scoring function is trained on FB15k-237, WN18RR, and YAGO3-10 and evaluated on held-out splits, so the MRR and Hits results are external empirical measurements. There are no self-citations used as load-bearing evidence: the polar-coordinate and sine-distance components are credited to HAKE, RotatE, and pRotatE, which are independent prior works, and no uniqueness theorem is imported. The skeptical concern about Eq. (4)'s boundary-continuity constant is a correctness and implementation issue about whether the soft score matches the hard membership semantics at sector boundaries; even if valid, it does not make the derivation circular, because the pattern proofs are stated at the level of the defined 'holds iff inside' semantics rather than derived from the score. Accordingly, no circular step can be exhibited, and the score is 0.

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

The central claim rests on the domain assumption that polar modulus encodes hierarchy, the simplifying equivalence between sector membership and triple truth, and the standard negative-sampling training objective. These are not derived from first principles.

free parameters (5)
  • beta (β)
    Introduced in the definition of sector central angle δ(r)=φ(r)-ψ(r)+βπ, controlling sector width; the paper lists grid ranges for other hyperparameters but not β, so the chosen value is unknown.
  • margin γ
    Fixed margin in loss, grid searched over {3,6,9,12,18,24,30}; best value not reported.
  • sampling temperature α
    Self-adversarial sampling temperature, grid searched over {0.5,1}; best value not reported.
  • embedding dimension d
    Chosen from {50,100,200,500,1000}; best value not reported.
  • λ1, λ2
    Learned scalar weights in scoring function (Eq. 5); final values not reported.
assumptions (3)
  • domain assumption Polar coordinate modulus represents semantic hierarchy.
    Central motivation borrowed from HAKE; no evidence provided that a single modulus component can encode the ordering animal > mammal > dog in a way that aids prediction.
  • domain assumption Triple truth is equivalent to bumped points lying inside the relation sector.
    Pattern proofs in Section III-D use strict set inclusion, but the scoring function is a soft distance, so the equivalence is an idealization.
  • standard math Negative sampling with self-adversarial temperature, adopted from RotatE, yields a valid training objective.
    Standard practice in KGC; no new justification needed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge Graph Embeddings with Representing Relations as Annular Sectors." pith.science (2026). https://pith.science/paper/GLLX44AF

@misc{pith2026250611099,
  author       = {Pith},
  title        = {Pith review of: Knowledge Graph Embeddings with Representing Relations as Annular Sectors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GLLX44AF}},
  note         = {Machine review of arXiv:2506.11099}
}
read the original abstract

Knowledge graphs (KGs), structured as multi-relational data of entities and relations, are vital for tasks like data analysis and recommendation systems. Knowledge graph completion (KGC), or link prediction, addresses incompleteness of KGs by inferring missing triples (h, r, t). It is vital for downstream applications. Region-based embedding models usually embed entities as points and relations as geometric regions to accomplish the task. Despite progress, these models often overlook semantic hierarchies inherent in entities. To solve this problem, we propose SectorE, a novel embedding model in polar coordinates. Relations are modeled as annular sectors, combining modulus and phase to capture inference patterns and relation attributes. Entities are embedded as points within these sectors, intuitively encoding hierarchical structure. Evaluated on FB15k-237, WN18RR, and YAGO3-10, SectorE achieves competitive performance against various kinds of models, demonstrating strengths in semantic modeling capability.

Figures

Figures reproduced from arXiv: 2506.11099 by the authors.

Figure 1
Figure 1. Simple illustration of SectorE. r h and r t are the region embeddings of the relation r. The position of the entity ei is determined by a basic point ei and a transformational bump bj . R d + × [0, 2π) d . Consequently, the ultimate embedding of the entity ei in a certain triple r(ei , ej ) is defined as: e r(ei,ej ) i = ei ⊙ bj (1) To be exact, the specific embedding e r(ei,ej ) i can be repre￾sented as follows: e … view at source ↗
Figure 2
Figure 2. Visualization of several entity embeddings from WN18RR dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 25 canonical work pages

  1. [1]

    Translating embeddings for modeling multi-relational data,

    A. Bordes, N. Usunier, A. Garc ´ıa-Dur´an, J. Weston, and O. Yakhnenko, “Translating embeddings for modeling multi-relational data,” in Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neu- ral Information Processing Systems 2013. Proceedings of a meeting held December 5-8, 2013, Lake Tahoe, Nevada, United States , 2013, pp....

    work page 2013

  2. [2]

    A three-way model for collective learning on multi-relational data

    M. Nickel, V . Tresp, H.-P. Kriegel et al. , “A three-way model for collective learning on multi-relational data.” in Icml, 2011, pp. 3 104 482–3 104 584

    work page 2011

  3. [3]

    RotatE: Knowl- edge graph embedding by relational rotation in complex space,

    Z. Sun, Z. Deng, J. Nie, and J. Tang, “RotatE: Knowl- edge graph embedding by relational rotation in complex space,” in 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019

    work page 2019

  4. [4]

    Learning hierarchy-aware knowledge graph embeddings for link prediction,

    Z. Zhang, J. Cai, Y . Zhang, and J. Wang, “Learning hierarchy-aware knowledge graph embeddings for link prediction,” in Proceedings of the AAAI conference on TABLE IV: Geometric mean area per dimension for all relation anunlar sectors in Y AGO3-10 after training. Relation Box V olume Head Box Tail Box actedIn 1.3432 0.4743 created 1.1743 0.6641 dealsWith ...

    work page 2020

  5. [5]

    Learning hierarchy-aware quaternion knowl- edge graph embeddings with representing relations as 3d rotations,

    J. Yang, X. Ying, Y . Shi, X. Tong, R. Wang, T. Chen, and B. Xing, “Learning hierarchy-aware quaternion knowl- edge graph embeddings with representing relations as 3d rotations,” in Proceedings of the 29th international conference on computational linguistics, 2022, pp. 2011– 2023

    work page 2022

  6. [6]

    Convolutional 2d knowledge graph embeddings,

    T. Dettmers, P. Minervini, P. Stenetorp, and S. Riedel, “Convolutional 2d knowledge graph embeddings,” in Proceedings of the AAAI conference on artificial intel- ligence, vol. 32, 2018

    work page 2018

  7. [7]

    Observed versus latent features for knowledge base and text inference,

    K. Toutanova and D. Chen, “Observed versus latent features for knowledge base and text inference,” in Pro- ceedings of the 3rd workshop on continuous vector space models and their compositionality , 2015, pp. 57–66

    work page 2015

  8. [8]

    Y AGO3: A knowledge base from multilingual wikipedias,

    F. Mahdisoltani, J. Biega, and F. M. Suchanek, “Y AGO3: A knowledge base from multilingual wikipedias,” in Seventh Biennial Conference on Innovative Data Systems Research, CIDR 2015, Asilomar, CA, USA, January 4-7, 2015, Online Proceedings . www.cidrdb.org, 2015

    work page 2015

Show all 28 references
  1. [9]

    Query2box: Reason- ing over knowledge graphs in vector space using box embeddings,

    H. Ren, W. Hu, and J. Leskovec, “Query2box: Reason- ing over knowledge graphs in vector space using box embeddings,” CoRR, vol. abs/2002.05969, 2020

    2002 arXiv

  2. [10]

    BoxE: A box embedding model for knowledge base completion,

    R. Abboud, ˙I. ˙I. Ceylan, T. Lukasiewicz, and T. Salva- tori, “BoxE: A box embedding model for knowledge base completion,” in Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 20...

    2020

  3. [11]

    Knowledge graph embedding by translating on hyperplanes,

    Z. Wang, J. Zhang, J. Feng, and Z. Chen, “Knowledge graph embedding by translating on hyperplanes,” in Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, July 27 -31, 2014, Qu ´ebec City, Qu´ebec, Canada. AAAI Press, 2014, pp. 1112–1119

    2014

  4. [12]

    Learning entity and relation embeddings for knowledge graph completion,

    Y . Lin, Z. Liu, M. Sun, Y . Liu, and X. Zhu, “Learning entity and relation embeddings for knowledge graph completion,” in Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, January 25-30, 2015, Austin, Texas, USA. AAAI Press, 2015, pp. 2181– 2187

    2015

  5. [13]

    Knowledge graph embedding via dynamic mapping matrix,

    G. Ji, S. He, L. Xu, K. Liu, and J. Zhao, “Knowledge graph embedding via dynamic mapping matrix,” in Pro- ceedings of the 53rd annual meeting of the association for computational linguistics and the 7th international joint conference on natural language processing (volume 1: L...

    2015

  6. [14]

    ExpressivE: A spatio- functional embedding for knowledge graph completion,

    A. Pavlovic and E. Sallinger, “ExpressivE: A spatio- functional embedding for knowledge graph completion,” in The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,

    2023

  7. [15]

    Capturing knowledge graphs and rules with octagon embeddings,

    V . Charpenay and S. Schockaert, “Capturing knowledge graphs and rules with octagon embeddings,” in Proceed- ings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI 2024, Jeju, South Korea, August 3-9, 2024 . ijcai.org, 2024, pp. 3289–3297

    2024

  8. [16]

    Em- bedding entities and relations for learning and inference in knowledge bases,

    B. Yang, W. Yih, X. He, J. Gao, and L. Deng, “Em- bedding entities and relations for learning and inference in knowledge bases,” in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, Y . Bengio ...

    2015

  9. [17]

    Complex embeddings for simple link pre- diction,

    T. Trouillon, J. Welbl, S. Riedel, ´E. Gaussier, and G. Bouchard, “Complex embeddings for simple link pre- diction,” in International conference on machine learn- ing. PMLR, 2016, pp. 2071–2080

    2016

  10. [18]

    SimplE embedding for link prediction in knowledge graphs,

    S. M. Kazemi and D. Poole, “SimplE embedding for link prediction in knowledge graphs,” in Advances in Neural Information Processing Systems 31: Annual Con- ference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montr ´eal, Canada, 2018, pp. 4289–4300

    2018

  11. [19]

    Dual quaternion knowledge graph embeddings,

    Z. Cao, Q. Xu, Z. Yang, X. Cao, and Q. Huang, “Dual quaternion knowledge graph embeddings,” in Proceed- ings of the AAAI conference on artificial intelligence , 2021, pp. 6894–6902

    2021

  12. [20]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Asso- ciation for Computational Linguistics: Human Language Technolo...

    2019

  13. [21]

    Convolutional 2d knowledge graph embeddings,

    T. Dettmers, P. Minervini, P. Stenetorp, and S. Riedel, “Convolutional 2d knowledge graph embeddings,” in Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Ap- plications of Artificial Intelligence (IAAI-18), and the 8t...

    2018

  14. [22]

    A novel embedding model for knowledge base completion based on convolutional neural network,

    D. Q. Nguyen, T. D. Nguyen, D. Q. Nguyen, and D. Phung, “A novel embedding model for knowledge base completion based on convolutional neural network,” in Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human La...

    2018

  15. [23]

    Modeling relational data with graph convolutional networks,

    M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, and M. Welling, “Modeling relational data with graph convolutional networks,” in The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedings 15 , 2018, pp. 593–607

    2018

  16. [24]

    Learning to exploit long- term relational dependencies in knowledge graphs,

    L. Guo, Z. Sun, and W. Hu, “Learning to exploit long- term relational dependencies in knowledge graphs,” in International conference on machine learning . PMLR, 2019, pp. 2505–2514

    2019

  17. [25]

    KG-BERT: BERT for knowledge graph completion,

    L. Yao, C. Mao, and Y . Luo, “KG-BERT: BERT for knowledge graph completion,” CoRR, vol. abs/1909.03193, 2019. [Online]. Available: http://arxiv.org/abs/1909.03193

    1909 arXiv

  18. [26]

    Structure-augmented text representation learning for efficient knowledge graph completion,

    B. Wang, T. Shen, G. Long, T. Zhou, Y . Wang, and Y . Chang, “Structure-augmented text representation learning for efficient knowledge graph completion,” in Proceedings of the Web Conference 2021 , 2021, pp. 1737–1748

    2021

  19. [27]

    KICGPT: Large language model with knowledge in context for knowledge graph completion,

    Y . Wei, Q. Huang, Y . Zhang, and J. Kwok, “KICGPT: Large language model with knowledge in context for knowledge graph completion,” in Findings of the As- sociation for Computational Linguistics: EMNLP 2023 . Singapore: Association for Computational Linguistics, Dec. 2023, pp....

    2023

  20. [2023]

    OpenReview.net, 2023

    2023

Pith tools

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