- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies partial-order (e.g., is-a hierarchy) embeddings based on hyperbolic entailment cones. It identifies two limitations of existing cone methods such as Hyperbolic Entailment Cones (HEC) and Shadow Cones: (1) geometric rigidity, since a cone's aperture is determined solely by the node's radial position and is therefore isotropic; and (2) optimization stagnation, since binary containment objectives assign zero energy, and hence zero gradient, to entailed pairs that already lie inside the cone.
The proposed Hyperbolic Gravitational Cone (HGC) replaces HEC's global aperture constant with a learnable anisotropic "mass" , where the directional factor is produced by a small Möbius-layer network, and the cone aperture follows in closed form. A finite-difference regularizer enforces radial monotonicity and convexity of the mass, under which the paper proves ray-wise transitivity. Training minimizes a "gravitational energy" within a max-margin loss: a softplus-gated angular term weighted by the parent's (stop-gradient) mass and by a hyperbolic distance, measured to the cone axis for positive pairs and to the cone boundary for negative pairs. The paper also argues that HEC and Shadow Cones are special cases of HGC.
Experiments evaluate hierarchy reconstruction, i.e., predicting held-out transitive-closure edges, on WordNet Mammal, WordNet Noun, and the Microsoft Concept Graph. Embedding dimensions range from 2 to 10, and the fraction of transitive edges included in training is varied. Baselines are box embeddings, HEC, and Shadow Cones variants. HGC reports the best F1 in most settings, e.g., 80.5 vs. 69.4 on Mammal (d = 2, 50%). The appendix adds ablations of the anisotropic mass and the energy function, as well as visualizations of the learned cones and the energy landscape.
Critical strengths
1. A simple and interpretable generalization of entailment cones. The paper clearly articulates two limitations of existing cone embeddings: isotropic apertures, and zero gradients for entailed pairs inside the cone. It addresses them with lightweight, well-motivated changes: a learnable, direction-dependent mass that yields a closed-form aperture, and a softplus-gated energy. HEC is recovered as a special case (α = 0), so the idea is easy to understand and could be dropped into existing cone-based models.
2. Clear gains in low-dimensional and low-supervision settings. HGC performs particularly well when capacity or supervision is limited. For example, it reaches 80.5 vs. 69.4 F1 on Mammal at d = 2 with 50% of transitive edges, and 55.3 vs. 52.6 on Noun at d = 10 when only the transitive reduction is used for training. HGC also trains from a simple uniform initialization without a separate pre-training stage, and uses a single formulation across datasets instead of requiring a choice among Shadow Cones variants.
Critical weaknesses
Significance of the problem. It is difficult to assess how important the two problems addressed in this paper are. The abstract motivates partial-order modeling by its role in knowledge graph reasoning and multimodal hierarchical alignment, yet the experiments contain no results in either setting. Could the authors provide evidence, e.g., statistics or results from prior work, that such hierarchical structure plays a key role in practical NLP tasks?
Experiments. Overall, the experimental evaluation is too thin:
-
Datasets and applications. The evaluation covers only hierarchy reconstruction on WordNet Mammal, WordNet Noun, and MCG, and the method is not validated in any real application. Since the abstract highlights knowledge graph reasoning and multimodal alignment, I would expect at least one experiment in such a setting.
-
No direct evidence that the two problems are solved. I agree that, in principle, the anisotropic mass and the gravitational energy address the fixed-aperture and optimization-stagnation problems. However, beyond end-to-end F1 (including the ablations in Appendix I, which are shown only as plots and are not discussed in the main text), the experiments do not measure the effect of either component directly. For the learnable directional factor that controls the cone aperture, what does it actually achieve compared with an isotropic aperture? For example, do broader concepts receive wider cones (e.g., a correlation between aperture and subtree size), and do both false positives and false negatives decrease? Precision and recall are not reported separately. Similarly, I could not find any measurement of optimization stagnation, such as training curves, gradient norms of positive pairs inside the cone, or how far inside the cone the positive pairs lie, for HEC versus HGC. The qualitative evidence in Figs. 7 and 8 relies on a few selected nodes.
-
Comparison with language models. Given the rapid progress of language models, a comparison with LM-based approaches would be informative, e.g., embeddings from a pretrained encoder such as BERT, or LM-based hypernymy prediction on the same datasets. Such models use concept names, which the current structure-only setting does not, so this would not be a like-for-like baseline. I also do not mean that the proposed method lacks value if a language model performs better. Rather, such a comparison would clarify where geometric embedding methods currently stand.
Presentation. The font size in the figures is too small to read comfortably.
Questions for authors
see weaknesses
The paper is interesting, and the proposed method is fairly novel. However, its main weakness is that the experimental evaluation is too thin. If the authors can provide additional relevant or key experiments, I would consider raising my score.
3 comments
Q3: Comparison with language models.
R3: While LMs like BERT leverage rich lexical priors from concept names, in contrast to our pure structure, they operate in unconstrained Euclidean feature spaces that inherently lack mathematical guarantees for asymmetry and transitivity, making them prone to structural inconsistency and hallucinations across deep subtrees. To clarify where geometric methods stand, we will evaluate BERT' s baseline capability on hypernymy prediction. This paradigm directly aligns with recent advances in multi-modal hierarchical learning, such as [2], which proved that pairing foundation models with hyperbolic entailment cones provides the critical asymmetric inductive bias needed to resolve cross-granularity alignment without symmetric confusion. Consequently, geometric methods and LMs are fundamentally complementary rather than mutually exclusive: standalone HGC achieves strict topological fidelity at a tiny fraction of an LM's parameter footprint, while our gravitational formulation simultaneously serves as a plug-and-play geometric regularizer to enforce global hierarchical consistency in modern foundation models.
We added a baseline [3] using frozen BERT-base embeddings and an order-sensitive classifier trained on the same Mammal relation splits. Both 2D and 5D HGC achieved higher test F1 than BERT at all five training percentages. The largest gap occurred when training used only basic edges: BERT achieved 14.02% precision and 55.50% recall (22.39 F1), while 5D HGC achieved 66.7% precision and 74.6% recall (70.4 F1).
| Variant | Metric | Mammal | ||||
|---|---|---|---|---|---|---|
| 0% | 10% | 25% | 50% | 90% | ||
| Bert | precision | 14.02 | 25.60 | 44.65 | 48.11 | 75.00 |
| HGC (d=2) | precision | 48.0 | 62.2 | 67.7 | 77.4 | 69.6 |
| HGC (d=5) | precision | 66.7 | 71.5 | 84.4 | 82.3 | 89.2 |
| Bert | recall | 55.50 | 60.28 | 45.93 | 48.80 | 53.11 |
| HGC (d=2) | recall | 62.7 | 85.2 | 82.3 | 83.7 | 86.6 |
| HGC (d=5) | recall | 74.6 | 89.0 | 93.3 | 97.6 | 90.4 |
| Bert | f1 | 22.39 | 35.94 | 45.28 | 48.45 | 62.18 |
| HGC (d=2) | f1 | 54.4 | 71.9 | 74.3 | 80.5 | 77.2 |
| HGC (d=5) | f1 | 70.4 | 79.3 | 88.6 | 89.3 | 89.8 |
[2] Avik Pal, Max van Spengler, Guido Maria D’Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. In International Conference on Learning Representations, 2025.
[3] Michael Hanna and David Mareček. 2021. Analyzing BERT’s Knowledge of Hypernymy via Prompting. In Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 275–282, Punta Cana, Dominican Republic. Association for Computational Linguistics.
I appreciate the authors’ response. I am inclined to maintain my original score.
We sincerely thank the reviewer for the thorough, insightful, and constructive feedback. Below, we address each of your concerns in detail.
Q1: Significance of the problem.
R1: Real-world conceptual knowledge is intrinsically structured as hierarchical DAGs, meaning that encoding partial-order constraints establishes the intrinsically correct problem formulation that aligns with the data’s true logic. Crucially, semantic heterogeneity is ubiquitous in practical NLP ontologies: concepts residing at identical hierarchical depths often govern radically disparate sub-branch breadths (e.g., in WordNet, organism spans >14,000 descendants whereas its sibling monotreme spans <10), a fundamental asymmetry that isotropic cones rigidly fail to capture. Extensive literature across fine-grained classification, knowledge graph reasoning, and multimodal alignment has proven that enforcing geometric containment is vital for eliminating symmetric retrieval fallacies and preserving zero-shot generalization across abstraction levels. Following established literature (HEC, Shadow Cones), we evaluated pure hierarchical reconstruction to isolate and benchmark this geometric inductive bias without confounding artifacts from pre-trained text encoders, and we will explicitly integrate these practical task statistics and connections into the revised manuscript.
Q2: No direct evidence that the two problems are solved.
R2: We thank the reviewer for this constructive critique. We completely agree that end-to-end F1 alone is insufficient to isolate the individual contributions. We provide a correlation between aperture and subtree size (whether the learned mass m(u,Ωu) and aperture ψ(u) meaningfully capture intrinsic concept breadth rather than merely memorizing hierarchical depth (Review 2, Q4)), and compared with BERT (with Precision and recall reported in Q3), and explicitly qualify the transitivity violated ratio across directions (Review 1, Q2).