Pith. sign in

REVIEW 3 major objections 4 minor 48 references

TaxoBell represents taxonomy concepts as Gaussian boxes—probabilistic boxes whose covariance expresses ambiguity—and reports state-of-the-art parent placement across five benchmark taxonomies.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 10:32 UTC pith:4UBIHKTU

load-bearing objection Solid, well-engineered model with consistent wins, but the headline percentages don't match Table 1 and the dummy-token adaptation of TaxoEnrich is the load-bearing assumption. the 3 major comments →

arxiv 2601.09633 v2 pith:4UBIHKTU submitted 2026-01-14 cs.CL

TaxoBell: Gaussian Box Embeddings for Self-Supervised Taxonomy Expansion

Sahil Mishra , Srinitish Srinivasan , Srikanta Bedathur , Tanmoy Chakraborty This is my paper
classification cs.CL
keywords taxonomy expansionGaussian box embeddingsenergy-based learningBhattacharyya coefficientKL divergenceself-supervisionhierarchical containmentuncertainty
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

TaxoBell tries to show that taxonomy expansion—placing a new concept under the right parent in a structured hierarchy—is better served by probabilistic boxes than by point vectors or hard boxes. The paper represents each concept as an axis-aligned box carrying a Gaussian density, so the mean gives semantic location and the covariance gives calibrated uncertainty. It trains these Gaussian boxes with two complementary energy functions: a symmetric overlap term (Bhattacharyya coefficient) for semantic similarity and an asymmetric containment term (KL divergence) for the directional 'is-a' relation. Across five benchmark taxonomies, the authors report that TaxoBell outperforms eight baselines by 19% in MRR and about 25% in Recall@k, with ablations showing each component is needed. If true, this gives a practical recipe for keeping real-world product and biomedical taxonomies current without manual curation.

Core claim

The paper's central claim is that translating between box geometry and Gaussian distributions fixes what hard boxes and point vectors get wrong in hierarchy learning. A concept is encoded by a language model, projected to a box (center and positive offsets), and then to a diagonal Gaussian whose mean is the box center and whose covariance is the squared offsets. Because the Gaussian softens the box boundary, overlap and containment become smooth, closed-form quantities instead of piecewise geometric tests, so gradients remain stable at the boundaries that break hard-box methods. Training uses a self-supervised contrastive objective mined from the seed taxonomy's own edges: positive parent–ch

What carries the argument

The central object is the Gaussian box: an axis-aligned hyperrectangle equipped with a multivariate Gaussian density, parameterized by a center vector (the mean) and an offset vector that becomes the diagonal covariance. Its work is to give box embeddings a smooth, probabilistic interior, so that 'is-a' containment is expressed as KL divergence from child to parent and semantic similarity as Bhattacharyya coefficient, both in closed form. The training objective combines these two energies with a reverse-KL coverage margin (parents must be broader than children) and volume regularization (minimum and maximum variance), preventing collapse and inflation that would otherwise let the model fake

Load-bearing premise

The claim rests on the assumption that adapting the two completion-style baseline models to the leaf-attachment task by substituting a dummy/sentinel token for the missing child preserves their scoring quality; if that placeholder misrepresents what those models need, the reported gains over the strongest baseline are overstated.

What would settle it

Take the two strongest completion-style baselines, feed each test query's true parent together with a real child of that parent (or the actual context the method was designed for) as the placeholder input, and compare their rankings against TaxoBell's on the same five benchmarks. If their MRR rises to TaxoBell's level or beyond once the sentinel is replaced, the headline improvement is an artifact of the adaptation, not the representation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, automated taxonomy expansion for e-commerce catalogs, search, and biomedical terminologies can be driven by self-supervision from the existing hierarchy alone, without human-labeled pairs.
  • The two-energy training recipe—symmetric overlap plus asymmetric containment with a coverage margin—generalizes the box-embedding idea beyond taxonomies to any domain with asymmetric relations.
  • Because the Gaussian can be converted back to a box at a chosen confidence level, practitioners get a tunable precision/recall dial: wider boxes include more candidate parents, narrower boxes give stricter attachment.
  • The ablations indicate that removing any single component (symmetric, asymmetric, divergence, or regularization) measurably degrades ranking, so the full combination is load-bearing rather than a nice-to-have.
  • Multi-parent taxonomies benefit most, suggesting the model can allocate probability mass to several valid parents rather than forcing a single attachment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same Gaussian-box machinery could be applied to knowledge-graph completion and ontology alignment, where symmetric similarity has long been known to underperform directional containment; the paper does not test those tasks but nothing in the mechanism is taxonomy-specific.
  • The reported 19% MRR gain may partly reflect better handling of ambiguous surface names; a testable extension would be to measure TaxoBell's margin over baselines on intentionally polysemous query sets, where covariance should matter most.
  • One implicit consequence is that the learned covariances could serve as a confidence score for downstream human review: attach automatically when the top parent's mass is well separated from the runner-up, and flag borderline cases for a curator.
  • The authors compare against completion-style baselines adapted with a placeholder child; a strict test of the SOTA claim would rerun those baselines in their native setting or with the true child context supplied, to confirm the sentinel does not handicap them.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. TaxoBell models taxonomy nodes as Gaussian boxes obtained by projecting BERT encodings into axis-aligned boxes whose diagonal offsets define Gaussian covariances. Training is self-supervised from a seed taxonomy and combines a symmetric Bhattacharyya overlap loss, an asymmetric KL alignment loss, a reverse-KL coverage term, and volume regularizers. At inference, candidate parents are ranked by either the Bhattacharyya coefficient or the negative KL divergence. The paper reports consistent improvements over seven structural and vector baselines on five taxonomy-expansion benchmarks and claims average gains of 19% in MRR and 25% in Recall@k.

Significance. If the empirical results are reproducible, TaxoBell is a solid contribution to taxonomy expansion: it provides a smooth alternative to hard box embeddings, an explicit uncertainty representation, a self-supervised objective with clearly ablated components, and a public code link. The ablations and the two inference variants are informative. However, the headline performance numbers are not supported by Table 1, and the adaptation of the two strongest baselines is inadequately validated, so the SOTA claim is currently overstated.

major comments (3)
  1. [Abstract, §1, Table 1] The abstract and introduction state that TaxoBell outperforms eight baselines by 19% in MRR, 25% in Recall@k, 43% in MR, and 21% in Hit@k. I cannot reproduce these numbers from Table 1. Taking the best baseline per dataset (TaxoEnrich, or STEAM on Food MRR) and the better TaxoBell variant per metric, the average relative gain is about 12% for MRR (5.4% on Science, 17.9% on Environment, 13.5% on WordNet, 8.0% on Food, 15.4% on MeSH), roughly 17% for R@1/R@5, and roughly 15% for Hit@1/Hit@5 on the two datasets reporting Hit@k. Only the MR improvement is close to the stated 43%. Moreover, Table 1 does not contain a Recall@k column, and §5.1.2 enumerates seven baselines, not eight. Please correct the headline numbers and specify the exact baseline/metric aggregation used; if the stated percentages come from a different comparison (e.g., average over all baselines), the protocol should be des
  2. [§5.1.2, Appendix C] The two strongest structural baselines, TaxoEnrich and TMN, are designed for taxonomy completion and score triples (parent, child, query). To use them for leaf attachment, the missing child is replaced with a dummy/sentinel token. The manuscript asserts that this 'preserves their original scoring function,' but no evidence is provided. If the sentinel is out-of-distribution for encoders trained on real children, their scores will degrade and TaxoBell's margin over the best baseline will be inflated. Since TaxoEnrich is the decisive baseline for the SOTA claim, please add a control experiment that quantifies the effect of the sentinel (e.g., evaluate TaxoEnrich/TMN with a real child on a subset where gold children are available, or retrain with sentinel tokens), or temper the comparative claims.
  3. [§5.2, Table 2] The statistical significance claim is not reproducible as reported. The per-metric p-values are not shown, the test used to obtain them is not named (paired bootstrap? Wilcoxon? t-test over the five seed runs?), and the metrics within a dataset are correlated, violating the independence assumption underlying Fisher's method. Please report the full test procedure and either use a multivariate test or account for the correlation among metrics.
minor comments (4)
  1. [§1, §5.1.2] The paper alternates between 'eight state-of-the-art baselines' (abstract, introduction) and a list of seven baselines (Section 5.1.2). Please make the count consistent.
  2. [Tables 3–6 vs Table 1] Ablation numbers for the full model do not match Table 1. For example, Science MRR is 59.41 in Table 4 and 58.50 in Table 6 but 58.2 in Table 1; Environment R@1 is 50.91 in Table 4 but 46.5 in Table 1. If the ablation tables are from a different run or configuration, this should be stated; if not, the numbers need reconciliation.
  3. [§4.4, Eq. (11)] Ldiverge explicitly enforces a parent-broader-than-child volume gap via the hinge on D_KL(N_p||N_c) versus C*(logVol(N_p)-logVol(N_c)). The later discussion of 'learned hierarchical containment' should therefore be framed as a loss-imposed property, not an emergent behavior of the representation.
  4. [§5 and Appendix D] The phrase 'calibrated uncertainty' is used repeatedly, but no calibration metric is reported. The 68–95–99.7 rule is invoked as an interpretability device; please add a quantitative calibration check or soften the claim.

Circularity Check

0 steps flagged

No significant circularity; SOTA claim rests on held-out evaluation.

full rationale

TaxoBell's central claim is an empirical benchmark claim: it is trained on seed-taxonomy edges (child-parent positives, neighborhood negatives) and evaluated on held-out query nodes against eight baselines, so the reported MR/MRR/Recall improvements are measured, not fitted. The model definition is self-contained: the box-to-Gaussian projection (mean = center, covariance = diag(offset^2)) is a reparameterization, and the Bhattacharyya/KL energies are standard closed-form divergences used both for training and as rankers; there is no hidden reuse of test labels in the loss. The Ldiverge term explicitly forces a parent-broader-than-child coverage margin (Eq. 11: Ldiverge = max(0, C*D_rep - D_KL(N_p||N_c))), so the ablation in Table 6 is a design validation, not a prediction of that margin from data; indeed Table 13 shows Ldiverge can decrease MRR on Environment, so the gain is not by construction. The dummy-token adaptation of TaxoEnrich/TMN (Appendix C) is a baseline-fairness concern, not a circular step. Self-citations [27,28,29] appear only in related work and metric conventions and are not load-bearing. No derivation reduces to its input, so no significant circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim is empirical, so the free parameters are primarily training hyperparameters. There are no newly invented physical entities. Several hyperparameter values are missing or inconsistent, which mainly affects reproducibility. The key domain assumptions are the Gaussian-box containment modeling and the dummy-token adaptation of completion baselines.

free parameters (6)
  • loss weights for Lasym, Lsym, volume regularization = 0.45, 0.45, 0.10 (Appendix A)
    Chosen by hand/tuning; the overall loss in Eq. 15 is written without these weights.
  • lambda (weight of Ldiverge in Lasym) = 0.3 (Appendix A) and ~0.5 (Appendix F)
    Tuned on Science and Environment; the two locations give inconsistent values.
  • C (scale in Ldiverge) = 1.5
    Tuned on Science and Environment; used for all datasets.
  • number of hard negatives per query = 50 (Science), 50 (Environment), 10 (WordNet), 20 (MeSH), 20 (Food)
    Dataset-specific hyperparameter, not justified by a principled rule.
  • embedding dimension d = varied 16-512 in Fig. 3; default not explicitly stated
    The implementation appendix does not state the default dimensionality used for Table 1.
  • delta (triplet margin), delta_var (min variance), M_var (variance ceiling) = not reported
    Mentioned in Eqs. 10, 13, and 14 but their values are omitted, so exact reproduction is impossible.
axioms (5)
  • standard math Closed-form Bhattacharyya distance (Eq. 4) and KL divergence (Eq. 5) are valid for diagonal Gaussians.
    Standard results used without proof; also used by prior Gaussian embedding work [38].
  • domain assumption A child concept can be represented as a Gaussian whose mean and diagonal covariance are derived from a box (Eq. 8), and KL(child||parent) is a valid containment score.
    This is the core modeling assumption; the paper does not prove that box containment corresponds to KL ordering.
  • domain assumption BERT [CLS] of surface name plus definition encodes enough semantics for unseen query nodes.
    All methods rely on this, but TaxoBell's projection head is trained on seed-taxonomy pairs, so the query node's semantics must transfer from the pretrained encoder.
  • domain assumption Sampling negatives from the local neighborhood (siblings, uncles, cousins, grandparents) and excluding ancestors produces no harmful false negatives.
    Negative candidates are structurally close to the true parent; if many are effectively valid parents in multi-parent taxonomies, the contrastive signal could be noisy.
  • ad hoc to paper The dummy-token adaptation of TaxoEnrich and TMN preserves their intended scoring function.
    Appendix C introduces this adaptation for the leaf-attachment task; it is load-bearing for the comparison against these two baselines.

pith-pipeline@v1.3.0-alltime-deepseek · 20514 in / 13511 out tokens · 134075 ms · 2026-08-03T10:32:12.105336+00:00 · methodology

0 comments
read the original abstract

Taxonomies form the backbone of structured knowledge representation across diverse domains, enabling applications such as e-commerce and semantic search. Yet, manual taxonomy expansion is labor-intensive and slow. Existing methods rely on point-based vector embeddings, which model symmetric similarity and thus struggle with the asymmetric relationships that are fundamental to taxonomies. Box embeddings offer a promising alternative by enabling containment and disjointness, but they face key issues: (i) unstable gradients at the intersection boundaries, (ii) no notion of semantic uncertainty, and (iii) limited capacity to represent polysemy or ambiguity. We address these shortcomings with TaxoBell, a Gaussian box embedding framework that translates between box geometries and multivariate Gaussian distributions, where means encode semantic location and covariances encode uncertainty. Energy-based optimization yields stable optimization, robust modeling of ambiguous concepts, and interpretable hierarchical reasoning. Extensive experiments on five benchmark datasets demonstrate that TaxoBell significantly outperforms eight state-of-the-art taxonomy expansion baselines by 19% in MRR and around 25% in Recall@k. We further demonstrate the advantages and pitfalls of TaxoBell with error analysis and ablation studies.

Figures

Figures reproduced from arXiv: 2601.09633 by Sahil Mishra, Srikanta Bedathur, Srinitish Srinivasan, Tanmoy Chakraborty.

Figure 1
Figure 1. Figure 1: Overview of taxonomy expansion and the contribu [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of TaxoBell. Entities are encoded with 𝑓𝜂 (.), mapped to axis-aligned boxes using 𝑓𝜓 (.), and then pro￾jected to Gaussian embeddings. Training optimizes two en￾ergies on the Gaussians – a symmetric overlap term (Bhat￾tacharyya Coefficient) and an asymmetric containment term (KL Divergence). For visualization, the boxes are shown in 2D, while the learned embeddings can be higher-dimensional. and 𝑄,… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of embedding dimensionality on perfor [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Case study of reconstructing a MeSH multi-parent [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Heatmap of sensitivity of the Ldiverge hyperparam￾eters on Hit@1 for Science and Environment. We sweep the scale 𝐶 (vertical axis) and the weight 𝜆 (horizontal axis) [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 3 linked inside Pith

  1. [1]

    Ben Athiwaratkun, Andrew Wilson, and Anima Anandkumar. 2018. Probabilistic FastText for Multi-Sense Word Embeddings. InProceedings of ACL. 1–11

    2018

  2. [2]

    Ben Athiwaratkun and Andrew Gordon Wilson. 2017. Multimodal word distri- butions.arXiv preprint arXiv:1704.08424(2017)

    Pith/arXiv arXiv 2017

  3. [3]

    Mohit Bansal, David Burkett, Gerard de Melo, and Dan Klein. 2014. Structured Learning for Taxonomy Induction with Belief Propagation. InProceedings of ACL. 1041–1051

    2014

  4. [4]

    Georgeta Bordea, Paul Buitelaar, Stefano Faralli, Roberto Navigli, et al . 2015. Semeval-2015 task 17: Taxonomy extraction evaluation (texeval). InProceedings of SemEval. 902–910

    2015

  5. [5]

    Georgeta Bordea, Els Lefever, and Paul Buitelaar. 2016. SemEval-2016 Task 13: Taxonomy Extraction Evaluation (TExEval-2). InProceedings of SemEval. 1081– 1091

    2016

  6. [6]

    Haw-Shiuan Chang, Ziyun Wang, Luke Vilnis, and Andrew McCallum. 2018. Dis- tributional Inclusion Vector Embedding for Unsupervised Hypernymy Detection. InProceedings of NAACL. 485–495

    2018

  7. [7]

    Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. 2021. Box Embeddings: An open-source library for representation learning using geometric structures. In Proceedings of EMNLP. 203–211

    2021

  8. [8]

    Shib Dasgupta, Michael Boratko, Dongxu Zhang, Luke Vilnis, Xiang Li, and Andrew McCallum. 2020. Improving Local Identifiability in Probabilistic Box Embeddings. InNeurIPS, Vol. 33. 182–192

    2020

  9. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of NAACL. 4171–4186

    2019

  10. [10]

    Shizhu He, Kang Liu, Guoliang Ji, and Jun Zhao. 2015. Learning to Represent Knowledge Graphs with Gaussian Embedding. InProceedings of CIKM. 623–632

    2015

  11. [11]

    Marti A. Hearst. 1992. Automatic Acquisition of Hyponyms from Large Text Corpora. InCOLING 1992 Volume 2: The 14th International Conference on Compu- tational Linguistics

    1992

  12. [12]

    Jiaxin Huang, Yiqing Xie, Yu Meng, Yunyi Zhang, and Jiawei Han. 2020. Corel: Seed-guided topical taxonomy construction by concept learning and relation transferring. InProceedings of KDD. 1928–1936

    2020

  13. [13]

    Tony Jebara, Risi Kondor, and Andrew Howard. 2004. Probability Product Kernels. J. Mach. Learn. Res.5 (2004), 819–844

    2004

  14. [14]

    Minhao Jiang, Xiangchen Song, Jieyu Zhang, and Jiawei Han. 2022. Taxoenrich: Self-supervised taxonomy completion via structure-semantic representations. In Proceedings of WWW. 925–934

    2022

  15. [15]

    Song Jiang, Qiyue Yao, Qifan Wang, and Yizhou Sun. 2023. A Single Vector Is Not Enough: Taxonomy Expansion via Box Embeddings. InWWW. 2467–2476

    2023

  16. [16]

    Giannis Karamanolakis, Jun Ma, and Xin Luna Dong. 2020. TXtract: Taxonomy- Aware Knowledge Extraction for Thousands of Product Categories. InProceedings of ACL. 8489–8502

    2020

  17. [17]

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, Vol. 1. Minneapolis, Minnesota, 2

    2019

  18. [18]

    Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, Fujie Huang, et al. 2006. A tutorial on energy-based learning.Predicting structured data1, 0 (2006)

    2006

  19. [19]

    Dongha Lee, Jiaming Shen, SeongKu Kang, Susik Yoon, Jiawei Han, and Hwanjo Yu. 2022. Taxocom: Topic taxonomy completion with hierarchical discovery of novel topic clusters. InProceedings of WWW. 2819–2829

    2022

  20. [20]

    Xiang Li, Luke Vilnis, Dongxu Zhang, Michael Boratko, and Andrew McCallum

  21. [21]

    Carolyn E Lipscomb. 2000. Medical subject headings (MeSH).Bulletin of the Medical Library Association(2000), 265

    2000

  22. [22]

    Zichen Liu, Hongyuan Xu, Yanlong Wen, Ning Jiang, Haiying Wu, and Xiaojie Yuan. 2021. TEMP: taxonomy expansion with dynamic margin loss through taxonomy-paths. InProceedings of EMNLP. 3854–3863

    2021

  23. [23]

    Xusheng Luo, Luxin Liu, Yonghua Yang, Le Bo, Yuanpeng Cao, Jinghang Wu, Qiang Li, Keping Yang, and Kenny Q Zhu. 2020. AliCoCo: Alibaba e-commerce cognitive concept net. InProceedings of SIGMOD. 313–327

    2020

  24. [24]

    Mahabal, Jiyun Luo, Rui Huang, Michael Ellsworth, and Rui Li

    A. Mahabal, Jiyun Luo, Rui Huang, Michael Ellsworth, and Rui Li. 2023. Producing Usable Taxonomies Cheaply and Rapidly at Pinterest Using Discovered Dynamic 𝜇-Topics.ArXivabs/2301.12520 (2023)

    Pith/arXiv arXiv 2023

  25. [25]

    Emaad Manzoor, Rui Li, Dhananjay Shrouty, and Jure Leskovec. 2020. Expanding taxonomies with implicit edge semantics. InProceedings of WWW. 2044–2054

    2020

  26. [26]

    Yuning Mao, Tong Zhao, Andrey Kan, Chenwei Zhang, Xin Luna Dong, Christos Faloutsos, and Jiawei Han. 2020. Octet: Online catalog taxonomy enrichment with self-supervision. InProceedings of KDD. 2247–2257

    2020

  27. [27]

    Sahil Mishra, Kumar Arjun, and Tanmoy Chakraborty. 2025. Rank, Chunk and Expand: Lineage-Oriented Reasoning for Taxonomy Expansion. InFindings of ACL. 12935–12953

    2025

  28. [28]

    Sahil Mishra, Avi Patni, Niladri Chatterjee, and Tanmoy Chakraborty. 2025. QuanTaxo: A Quantum Approach to Self-Supervised Taxonomy Expansion.arXiv preprint arXiv:2501.14011(2025)

    arXiv 2025

  29. [29]

    Sahil Mishra, Ujjwal Sudev, and Tanmoy Chakraborty. 2024. FLAME: Self- Supervised Low-Resource Taxonomy Expansion Using Large Language Models. ACM Trans. Intell. Syst. Technol.(Dec. 2024)

    2024

  30. [30]

    Alexander Panchenko, Stefano Faralli, Eugen Ruppert, Steffen Remus, Hubert Naets, Cédrick Fairon, Simone Paolo Ponzetto, and Chris Biemann. 2016. A Taxonomy Induction Method based on Lexico-Syntactic Patterns, Substrings and Focused Crawling. InProceedings of SemEval. 1320–1327

    2016

  31. [31]

    Dhruvesh Patel, Shib Sankar Dasgupta, Michael Boratko, Xiang Li, Luke Vil- nis, and Andrew McCallum. 2020. Representing Joint Hierarchies with Box Embeddings. InAutomated Knowledge Base Construction

    2020

  32. [32]

    Hongyu Ren, Weihua Hu, and Jure Leskovec. 2020. Query2box: Reasoning over Knowledge Graphs in Vector Space Using Box Embeddings. InICLR

    2020

  33. [33]

    Jiaming Shen, Zhihong Shen, Chenyan Xiong, Chi Wang, Kuansan Wang, and Jiawei Han. 2020. TaxoExpan: Self-supervised taxonomy expansion with position- enhanced graph neural network. InProceedings of WWW. 486–497

    2020

  34. [34]

    Jiaming Shen, Zeqiu Wu, Dongming Lei, Chao Zhang, Xiang Ren, Michelle T Vanni, Brian M Sadler, and Jiawei Han. 2018. Hiexpan: Task-guided taxonomy construction by hierarchical tree expansion. InProceedings of KDD. 2180–2189

    2018

  35. [35]

    Rion Snow, Daniel Jurafsky, and Andrew Ng. 2004. Learning Syntactic Patterns for Automatic Hypernym Discovery. InNeurIPS, Vol. 17

    2004

  36. [36]

    Nikhita Vedula, Patrick K Nicholson, Deepak Ajwani, Sourav Dutta, Alessandra Sala, and Srinivasan Parthasarathy. 2018. Enriching taxonomies with functional domain knowledge. InSIGIR. 745–754

    2018

  37. [37]

    Luke Vilnis, Xiang Li, Shikhar Murty, and Andrew McCallum. 2018. Probabilistic Embedding of Knowledge Graphs with Box Lattice Measures. InProceedings of ACL. 263–272

    2018

  38. [38]

    Luke Vilnis and Andrew McCallum. 2014. Word representations via gaussian embedding.arXiv preprint arXiv:1412.6623(2014)

    Pith/arXiv arXiv 2014

  39. [39]

    Suyuchen Wang, Ruihui Zhao, Xi Chen, Yefeng Zheng, and Bang Liu. 2021. Enquire one’s parent and child before decision: Fully exploit hierarchical structure for self-supervised taxonomy expansion. InProceedings of WWW. 3291–3304

    2021

  40. [40]

    Suyuchen Wang, Ruihui Zhao, Yefeng Zheng, and Bang Liu. 2022. Qen: Applicable taxonomy completion via evaluating full taxonomic relations. InProceedings of WWW. 1008–1017

    2022

  41. [41]

    Hongyuan Xu, Yunong Chen, Zichen Liu, Yanlong Wen, and Xiaojie Yuan. 2022. TaxoPrompt: A Prompt-based Generation Method with Taxonomic Context for Self-Supervised Taxonomy Expansion. InProceedings of IJCAI. 4432–4438. Main Track

    2022

  42. [42]

    Hongyuan Xu, Ciyi Liu, Yuhang Niu, Yunong Chen, Xiangrui Cai, Yanlong Wen, and Xiaojie Yuan. 2023. TacoPrompt: A Collaborative Multi-Task Prompt Learning Method for Self-Supervised Taxonomy Completion. InProceedings of EMNLP. 15804–15817

    2023

  43. [43]

    Wei Xue, Yongliang Shen, Wenqi Ren, Jietian Guo, Shiliang Pu, and Weiming Lu

  44. [44]

    Yue Yu, Yinghao Li, Jiaming Shen, Hao Feng, Jimeng Sun, and Chao Zhang. 2020. Steam: Self-supervised taxonomy expansion with mini-paths. InProceedings of KDD. 1026–1035

    2020

  45. [45]

    Jieyu Zhang, Xiangchen Song, Ying Zeng, Jiaze Chen, Jiaming Shen, Yuning Mao, and Lei Li. 2021. Taxonomy completion via triplet matching network. In Proceedings of AAAI, Vol. 35. 4662–4670. WWW ’26, April 13-17, 2026, Dubai, United Arab Emirates Mishra et al. Table 9: Statistics of benchmark datasets.|N 0| and|E 0| denote the number of nodes and edges in ...

    2021

  46. [46]

    transport

    Yuchen Zhang, Amr Ahmed, Vanja Josifovski, and Alexander Smola. 2014. Tax- onomy discovery for personalized recommendation. InProceedings of WSDM. 243–252. Appendix A Implementation Details TaxoBellis implemented using PyTorch with the baselines taken from the respective repositories of their original authors. All train- ing and inference tasks were condu...

    2014

  47. [2019]

    Smoothing the Geometry of Probabilistic Box Embeddings. InICLR

  48. [2024]

    InProceedings of ACL

    Insert or Attach: Taxonomy Completion via Box Embedding. InProceedings of ACL. 3851–3863