acceptodds
Under review as a conference paper at ICLR 2027

MRNBOX: Ripple Diffusion Meets Geometric Reasoning for Taxonomy Completion

Abstract

The Taxonomy Completion (TC) task aims to ffnd optimal insertion positions for new concepts under hypernym-hyponym constraints, but faces three challenges: sparse descriptions and polysemy weaken semantic embeddings with no quantiffable parent-priority mechanism; traditional GNNs miss path-dependent decay in hierarchical correlations; and global traversal is inefffcient with invalid candidate interference. To address these, we propose MRNBOX, which resolves semantic sparsity via adaptive embeddings fusing conceptual descriptions, average path lengths, and topological association features; models hierarchical structures via ripple diffusion and sparse Mixture-of-Experts (MoE) for dynamic neighbor screening; and performs two-stage geometric reasoning, where box embedding exclusion ffrst prunes irrelevant branches before ffne-grained matching reffnes positioning.This enables efffcient semantic-structural collaborative modeling for dynamic taxonomy maintenance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Taxonomy Completion, Box embedding, Ripple diffusion, GNN

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_e549OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors submitted a paper titled “MRNBOX: Ripple Diffusion Meets Geometric Reasoning for Taxonomy Completion”

The submission includes a main PDF

The paper is organized in a standard structure. The required AI USE STATEMENT and the recommended REPRODUCIBILITY STATEMENT are both not included. An appendix is also included.

The paper addresses a problem related to the Taxonomy Completion task and proposes MRNBOX, which overcomes semantic sparsity by fusing conceptual descriptions, average path lengths, and topological features into adaptive embeddings. It models hierarchies via rippled diffusion and a sparse Mixture-of-Experts (MoE) for dynamic neighbor filtering, executing a two-stage geometric reasoning process that uses box embedding exclusion to prune irrelevant branches prior to fine-grained matching.

Experiments are conducted on four datasets: MAGCS(MC), MAGPSY(MP) WordNet-Verb, SemEval-Food( and against the following systems: TMN, QEN, TaxoEnrich, TaxoComplete, TaxBox, and CoSTC.

Critical strengths

The paper addresses an interesting topic for the community. The approach is formalized.

Critical weaknesses

The paper has some critical weaknesses:

  1. The required AI use statement is missing
  2. The recommended REPRODUCIBILITY STATEMENT is missing
  3. The overall presentation is compromised by: 3.a) There are several references that are not included: jurgens2016semeval bordea2015semeval zhang2021taxonomy wang2022qen jiang2022taxoenrich arous2023taxocomplete xue2023insert niu2024contrastive 3.b) Some layout problems, for instance, Figure 4 overlaps the text of Section 4.4
  4. Overall, there is room for improvement in different aspects of the manuscript
Questions for authors

Why does the adopted protocol not consider the general structural measures proposed in the SemEval tasks?

Reviewer_MbKoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses taxonomy completion (TC): inserting a new concept as a ⟨parent, child⟩ pair (non-leaf) or under a parent alone (leaf). MRNBOX has three modules. (a) A node embedding that concatenates a CLIP text embedding of the concept description, a CLIP embedding of templated sentences about the node's ancestors, descendants and siblings, and a 3-d vector [#ancestors, #descendants, mean distance to ancestors]. (b) "Ripple sets": k-hop BFS neighborhoods (siblings added at hop 1), with five scoring functions called "experts" (Child, Sibling, Parent for hops ≤ 2; Cross-Tree with subtree Jaccard and Hub with degree × betweenness for hops ≥ 3), mixed by a softmax gate that depends on the layer and then used to weight neighbor aggregation. (c) The aggregated embedding goes through a Highway network into a box. At test time the query box is compared with the boxes of the root's direct children. Only the top-k subtrees are kept, and candidate positions inside them are ranked by a product of box similarities. Experiments on SemEval-Food, WordNet-Verb, MAG-CS and MAG-PSY report large gains over TMN, QEN, TaxoEnrich, TaxBox and CoSTC, along with ablations, a leaf/non-leaf breakdown, sensitivity to top-k, box dimension and ripple depth, and a case study.

Critical strengths
  1. The motivation for coarse-to-fine pruning is reasonable. Using box containment to drop whole top-level subtrees before fine-grained matching could cut the cost of scoring every candidate position, which matters for large taxonomies such as MAG-CS (~150k candidates). If this were properly validated, it would be a useful practical contribution.

  2. The evaluation covers the standard TC benchmarks. All four common datasets are used, and the paper reports a leaf/non-leaf breakdown, which makes it easy to compare against prior work.

Critical weaknesses
  1. MRNBOX reports inconsistent results across tables (e.g., MAG-PSY MR = 212 vs. 199). Some metrics are also questionable (e.g., P@1 > Hit@1), and several reported improvements are incorrectly calculated.

  2. The ablation results do not always support the claims. For example, removing L2 improves performance on several metrics.

  3. Important training details, including the loss function and negative sampling strategy, are missing. The use of structural features for unseen query nodes is unclear and raises concerns about possible data leakage.

  4. The proposed sparse MoE and ripple diffusion appear similar to existing methods, and their novelty is unclear. Several equations also appear incorrect.

  5. The hierarchical pruning method lacks evidence for its reliability and efficiency. Pruning recall and runtime comparisons are needed.

  6. Several citations are unresolved, relevant baselines are missing, and the choice of CLIP's text encoder is not well justified.

Questions for authors
  1. Please explain the discrepancies between Tables 2 and 3 and provide precise definitions of P@k, R@k, and Hit@k. How can P@1 exceed Hit@1 or R@10 exceed Hit@10? Were all baselines evaluated using the same code and protocol?

  2. What is the complete training objective, including loss functions, negative sampling, and trainable parameters? How are structural features computed for unseen query nodes without known ancestors or descendants, and how is information leakage prevented?

  3. What guarantees the correctness of hierarchical pruning when learned boxes may violate parent-child containment? Please report pruning recall, runtime, and memory costs compared with CoSTC and TaxBox, including preprocessing overhead.

  4. Why does removing L2 improve performance on MAG-CS, despite the claimed complementary contributions of all modules? Please report results across multiple seeds and clarify whether hyperparameters were selected using validation or test data.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.