acceptodds
Under review as a conference paper at ICLR 2027

Harnessing Structural Context for Entity Alignment Foundation Models

Abstract

Entity alignment (EA) across unseen knowledge graph (KG) pairs requires a model to reuse alignment knowledge without learning pair-specific entity embeddings. In a structure-only setting, this raises two connected challenges: evidence from the two KGs must interact through limited seed anchors, and structurally similar candidates must be distinguished beyond coarse similarity. We propose ContextEA, a transferable EA framework that learns cross-KG context construction and candidate discrimination jointly. For each query, its encoder propagates query-relative, relation-conditioned structural signals over an anchor-bridged unified graph. Its decoder then uses entity, neighborhood, relation, and anchor evidence to calibrate ambiguous candidates; the calibration loss also updates the encoder during training. This design enables a pretrained model to apply learned structural alignment patterns to new KG pairs. Across 29 benchmark settings from OpenEA, SRPRS, and DBP, ContextEAimproves over structure-only transferable EA baselines. Without target-specific finetuning, it also exceeds the finetuned EAFM baseline across all three benchmark groups.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Entity Alignment, Foundation Models, Graph Neural Networks, Structural Context, Knowledge Graphs

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Ei9pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ContextEA, a structure-only framework for transferable entity alignment across knowledge graph pairs. It extends EAFM with query-conditioned cross-KG entity propagation through seed-alignment bridges and a structural calibration decoder that reranks top candidates using entity, neighborhood, relation, and anchor evidence. Experiments on 29 benchmark settings show consistent improvements over existing transferable baselines, though transfer uses seed alignments from each target pair.

Critical strengths
  1. The paper addresses a relevant challenge: transferring entity alignment models to new knowledge graph pairs using structure alone.
  2. ContextEA combines cross-KG propagation with candidate calibration, and reports consistent gains over EAFM and ULTRA-EA on the evaluated benchmarks.
  3. The paper includes useful ablations, efficiency measurements, and analyses of anchor robustness and failure cases.
Critical weaknesses
  1. The methodological contribution is not clearly distinguished from EAFM. For example, the fifth paragraph of the Introduction appears to describe the proposed method and EAFM’s processing steps together, making it difficult to identify which components are inherited from EAFM and which are introduced by this paper.

  2. The framework’s generality is unclear. ContextEA is built entirely on EAFM, so it is uncertain whether the proposed components can be applied to other transferable entity alignment methods. The paper should clarify whether the approach is specific to EAFM or can generalize beyond it.

  3. The experimental comparisons are limited. The evaluation compares ContextEA with only two methods. A broader comparison with relevant transferable entity alignment approaches is needed to better establish its performance relative to the field.

Questions for authors

See the Weaknesses above

Reviewer_DU6zOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ContextEA, a transferable entity alignment framework that enhances cross-KG structural interaction. The framework learns cross-KG context construction and candidate discrimination jointly. The authors evaluate ContextEA on 29 benchmark settings and report improvements over existing baselines.

Critical strengths

S1. The paper investigates transferable entity alignment, a good research problem.

S2. The proposed framework is straightforward and clearly presented.

Critical weaknesses

W1. Limited methodological novelty. The proposed approach mainly combines anchor-based cross-KG message passing with conventional multi-view structural feature fusion. Both of them are straightforward extensions of existing EA techniques. The paper does not sufficiently establish what new capability is introduced.

W2. Insufficient ablation studies. Table 3 shows that removing the decoder only drops the average MRR from 0.669 to 0.653, suggesting that most of the improvements come from the encoder. However, the authors didn’t compare it with a simple EAFM variant that just adds cross-KG anchor edges. So, it’s still unclear whether these significant gains are due to a simple architecture change.

W3. Concerns about experimental reliability. Tables 6 and 7 show that ULTRA-EA improved MRR by exactly 0.030 on 23 of the 24 datasets after fine-tuning. Such a uniform improvement is rare in heterogeneous datasets and needs explanation. In addition, the experimental values in Figure 4 are inconsistent with Tables 1 and 9. These discrepancies weaken confidence in the reported results and should be addressed.

Questions for authors

Q1. What is the main methodological contribution beyond adding cross-KG anchor edges and combining existing structural features?

Q2. Can the authors explain why finetuning consistently improves ULTRA-EA's MRR by exactly 0.030 across 23 of 24 datasets in Tables 6 and 7?

Q3. Can the authors clarify the discrepancies between Figure 4 and Tables 1 and 9?

Reviewer_FBr2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ContextEA, a transferable entity alignment (EA) model that uses only graph structure and builds on the EA foundation model EAFM. It keeps several EAFM components unchanged:

  • the query-relative merged relation graph;
  • the anchor-conditioned initialization;
  • the coarse matcher;
  • the full-softmax cross-entropy loss. It changes two things:
  • Encoder. The two KGs are joined into one entity graph. Seed alignments become bidirectional "bridge" edges with a learnable relation type. EAFM's softmax neighbor attention is replaced by independent sigmoid gates per edge, with sum aggregation and LayerNorm residuals.
  • Decoder. The top-16 candidates from the coarse matcher are re-scored with a bounded residual, β·tanh(f_dec(φ)). The features φ are entity, neighbor, relation-consistency and anchor-pair features. This part is trained with an extra loss over the gold candidate plus hard negatives.

Model is trained on D-W-15K-V1 and EN-DE-15K-V1 and evaluated on 29 OpenEA/SRPRS/DBP15K datasets in two ways: pretrain (no target training) and after fine-tuning on the target. The pretrain model beats fine-tuned EAFM in MRR on all 29 datasets. The ablation shows that most of the gain comes from the encoder and a smaller part from the decoder .

Critical strengths
  • Strong, consistent result in a hard setting. A single structure-only model, with no training on the target, beats a fine-tuned foundation-model baseline in MRR on all 29 datasets, including the 100K-scale ones.
  • Simple, well-specified design. The decoder is bounded and only touches the top-k, so it is cheap and cannot arbitrarily override the coarse ranking. Masking the supervised pair's own bridge during training prevents the most obvious leakage.
  • Thorough diagnostics: multiple seeds, robustness to anchor ratio and anchor noise, query buckets by anchor distance, failure analysis, and explicit feature definitions.
Critical weaknesses

W1. The contribution is incremental over EAFM, and the paper doesn't present the difference as such.

  • Inherited from EAFM unchanged: the coarse scoring function, the coarse loss, the merged relation graph, and the query-relative/anchor-conditioned initialization.
  • New, but each is close to existing work:
    • Connecting the two KGs through seed alignments is a common idea in earlier EA work (e.g., sameAs links in EAFM, parameter sharing in IPTransE).
    • Reranking the top-k candidates with neighbor and anchor features is similar to neighborhood-matching methods (e.g., NMN, RNM, BERT-INT), which are not discussed. Incremental work is acceptable if the paper explains why each change matters. Right now it mainly shows that the changes help.

W2. The direct contradiction with EAFM is not addressed.

  • EAFM's ablation: replacing parallel per-KG encoding with propagation across KGs drops average MRR from 0.529 to 0.447. EAFM attributes this to a "reasoning horizon gap" and states that parallel encoding is beneficial because it limits information flow between the two KGs.
  • EAFM's variant that replaces the merged relation graph with sameAs bridges also drops (0.484).
  • ContextEA brings cross-KG propagation back through bridges and reports +0.124 MRR.
  • The encoder also replaces softmax attention with sigmoid gates at the same time as adding bridges. The ablation removes both together, so we can't tell whether the gain comes from the bridges or from the new gating. "Structural context helps" cannot explain both outcomes. Something specific must make it work here: the gating, keeping the anchor-conditioned initialization, the learnable bridge relation, the bridge masking, or building one unified graph instead of propagating from the query. Without isolating that, we can't be sure of the reason for the gain or of the actual contribution.

W3. The decoder's role is not well supported.

  • The decoder adds +0.016 MRR. Combined with EAFM's encoder it lowers MRR (0.501 vs 0.529), so it is not a general-purpose reranker. Its benefit depends on ContextEA's encoder, and the paper doesn't explain why.
  • Training and inference don't match. In training, the decoder sees the gold candidate plus 5 hard negatives. At test time it reranks the top-16 candidates, where the gold may not be present at all. The paper doesn't discuss how this gap affects the decoder.

W4. Design and hyperparameter choices lack justification. Several key values are given without explanation or analysis:

  • β = 0.3: the paper says it follows EAFM, but EAFM has no decoder or β. It also seems to have been picked by looking at test datasets.
  • λ_dec = 0.2.
  • The decoder starts at epoch 20.
  • 5 hard negatives.
  • A cap of 24 neighbors in the anchor view.
  • k = 16, even though Table 9 shows k = 32 works better. A short explanation of how each was chosen would help.
Questions for authors
  1. EAFM contradiction. EAFM reports that propagation across KGs lowers MRR (0.529 → 0.447). Which ContextEA component makes it beneficial instead? Can you run EAFM's "w/o parallel encoding" variant with your gating, or ContextEA with softmax attention?

  2. Gate vs. softmax. Why did you replace softmax attention with sigmoid gates? Could you report EAFM with bridges but softmax kept, and ContextEA with gates but no bridges, to show which change drives the gain?

  3. Hyperparameters. How were β, λ_dec, the epoch-20 start, the number of negatives, the 24-neighbor cap and k chosen? Were they selected on validation data?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.