Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases

Summary

The paper proposes to match entities in dangling settings, where the entities may not have a link to any other entity. New dataset GA16K is proposed.

Strengths

The task is important in the knowledge graph field. The method achieves better F1 in the experiments. There are proofs for the correctness of the algorithm.

Weaknesses

Writting should be improved. For example, there is no definition of 'GA16K' in the whole paper. The first clear definition of 'GA16K' is in the appendix. Baselines are old. Many related works are missing (https://arxiv.org/abs/2210.10436, https://aclanthology.org/2022.acl-long.405/, https://dl.acm.org/doi/abs/10.1145/3404835.3462870, https://aclanthology.org/2021.emnlp-main.226/) The GA16K dataset seems to be multi-modal, but there are well-established datasets for multi-modal entity matching (https://dl.acm.org/doi/10.1145/3534678.3539244, https://arxiv.org/abs/2212.14454) that are not mentioned or used in the experiments. No large scale experiments are conducted. The paper should conduct experiments on large-scale datasets (https://arxiv.org/abs/2108.05211) to show the scalability of the proposed method.

Questions

See above

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

No limitations section is provided.

Summary

The paper tries to tackle the challenge of entity alignment (EA) with unlabeled dangling cases in knowledge graphs (KGs), where some entities lack counterparts in another KG. It presents a framework to detect dangling entities and align matchable entities using a GNN-based encoder and a positive-unlabeled learning algorithm, respectively. It also provides theoretical guarantees for the proposed methods, including unbiasedness, uniform deviation bounds, and convergence. Experimental results demonstrate the promising performance of the framework over baselines on real-world datasets, even without labeled dangling entities for training.

Strengths

- The paper addresses the challenge of entity alignment in a dangling-aware context, even when labeled data for dangling entities is unavailable. This practical scenario is explored, and can inspire future work on this task. - In my opinion, the proposed method, selective neighborhood aggregation combined with positive-unlabeled learning, holds promise for tackling the problem under investigation. - Experimental results across multiple datasets demonstrate that the proposed method can outperform baselines for dangling entity detection and entity alignment in most metrics.

Weaknesses

- The discussion on related work appears insufficient. I recommend enhancing the appendix by providing a detailed discussion and comparison. For instance, although the proposed GNN looks similar to Dual-AMN, no explicit discussions regarding this similarity are currently included. - In Table 2, the proposed GNN exhibits significant superiority over MTransE and AliNet. However, in Table 4, the advantage is reduced. Although this could be attributed to various factors, such as dataset characteristics, evaluation metrics, or specific scenarios, it is essential to carefully analyze the experimental setup and consider potential confounding variables to fully understand this discrepancy.

Questions

- According to Table 2, the proposed GNN greatly outperforms MTransE and AliNet. Why does it not show such a huge advantage in Table 4? - (Open question) Is the proposed method applicable to dangling entity detection with labeled data? What about its performance in this case?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

NA.

Summary

This paper elaborates the unique challenges of unlabeled dangling entities in EA task. To address the challenge, it proposes the framework, namely Lambda, for dangling detection and then achieves entity alignment. The main idea is to perform selective aggregation with spectral contrastive learning and to adopt theoretically guaranteed PU learning to relieve the dependence on the labeled dangling entities. Extensive experiments demonstrate the effectiveness of each module and its ability to handle Unlabeled Dangling Cases.

Strengths

1. The paper is well-motivated. 2. The paper propose innovative methods to addresses the challenge of unlabeled dangling entities in entity alignment tasks 3. Extensive experiments demonstrate the model's effectiveness, and the experiments unaware(aware) of dangling entities is particularly insightful. 4. The paper is well written and organized.

Weaknesses

1. The section on KG Entity Encoder with Selective Aggregation section adds an adaptive dangling indicator based on the Dual-AMN[1] model. The paper would be better to provide a detailed explanation of why the adaptive dangling indicator is effective. 2. In some evaluation metrics of datasets, Lamada's results still need improvement. For instance, the results for H@10 and H@50 in Table 2 are lower than those of Dual-AMN[1], but H@1 is more higher. And the precision in Tables 3 and 4 is slightly inferior. 3. Comparing the proposed method with strong baseline models under different ratios of pre-aligned seeds would better demonstrate the method's superiority. [1] Xin Mao, Wenting Wang, et al. Dual attention matching network with normalized hard sample mining. WWW 2021

Questions

1. The paper mention that the initialization of r_{e_j} is critical (L94-95). What are the potential consequences of poor initialization? How is a good initialization chosen? Is it based on experience or theoretical support? 2. How is the equivalent form of infoNCE (L137-137) derived? 3. What point of aligned entity sparsity does the model stop all subsequent processes and determine that the two KGs cannot be aligned? Is there experimental evidence supporting this?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

NA

Summary

This paper introduces a novel entity alignment framework called Lambda for aligning entities with dangling cases. It includes a GNN encoder, KEESA, to aggregate information within and across KGs, and an iterative positive-unlabeled learning algorithm, iPULE, to detect dangling entities. The authors provide both theoretical proof and empirical evidence to demonstrate the superiority of the proposed method.

Strengths

The idea of using positive pairs to support unlabeled dangling detection is interesting and effective. Results on dangling entity detection are promising. Theoretical proofs are provided to further support the proposed method.

Weaknesses

The writing may be improved. Figures need more detailed captions. The methodology section introduces too many new terms. For instance, in Lemma 1, perhaps using e_i, e_+ and e_j is better than q, p^+, p^-. The overall subscripts and superscripts are also inconsistent. The motivation of this paper is not clear enough. There are already many methods leveraging inter-graph and cross-graph GNNs for entity alignment. In this sense, the novelty of KEESA is limited. The so-called spectral contrastive learning seems no different from the existing ones. Then, the core contribution of this paper is iPULE. This paper could be improved if the authors delve deeper into the discussion of iPULE instead of KEESA, especially regarding the application of this module on different EA methods.

Questions

Why do the authors call L_info ``spectral contrastive learning''?

Rating

5

Confidence

4

Soundness

2

Presentation

1

Contribution

3

Limitations

N/A