Source-linked AI summary
Multi-view Knowledge Graph Embedding for Entity Alignment
Qingheng Zhang, Zequn Sun, Wei Hu, Muhao Chen, Lingbing Guo, Yuzhong Qu
TL;DR
Embedding-based entity alignment has not fully exploited the diverse features of knowledge-graph entities, with existing methods often using only one or two feature types. MultiKE learns embeddings from name, relation, and attribute views, combines them with several strategies, and uses cross-KG inference; experiments show it outperforms existing embedding-based methods across two real-world datasets.
Problem
Existing embedding-based entity alignment methods often exploit only one or two entity feature types, although different features characterize complementary aspects of entity identities.
Method
MultiKE divides entity features into complementary name, relation, and attribute views, learns view-specific embeddings, jointly combines them, and applies cross-KG inference methods.
Results
MultiKE largely outperforms existing embedding-based entity alignment methods on two real-world datasets, with selected views, cross-KG inference, and combination strategies all contributing to improvement.
Takeaways & Limitations
MultiKE provides comprehensive entity embeddings for alignment and achieves promising unsupervised results comparable to conventional entity alignment methods.
Takeaways & Limitations
The shared-space combination assumes that multiple views can be generated from a shared latent view.
Abstract
from arXiv · showhide
We study the problem of embedding-based entity alignment between knowledge graphs (KGs). Previous works mainly focus on the relational structure of entities. Some further incorporate another type of features, such as attributes, for refinement. However, a vast of entity features are still unexplored or not equally treated together, which impairs the accuracy and robustness of embedding-based entity alignment. In this paper, we propose a novel framework that unifies multiple views of entities to learn embeddings for entity alignment. Specifically, we embed entities based on the views of entity names, relations and attributes, with several combination strategies. Furthermore, we design some cross-KG inference methods to enhance the alignment between two KGs. Our experiments on real-world datasets show that the proposed framework significantly outperforms the state-of-the-art embedding-based entity alignment methods. The selected views, cross-KG inference and combination strategies all contribute to the performance improvement.
1 Introduction
Embedding-based entity alignment is limited by incomplete use of entity features and reliance on costly seed alignments. MultiKE addresses these issues by jointly embedding complementary name, relation, and attribute views, adding cross-KG inference and combination strategies.
- Entity alignment matches entities in different knowledge graphs that refer to the same real-world identity and supports KG construction, fusion, search, question answering, and recommendation.
- Existing embedding methods typically use only relational structure or add one supplementary feature type, although names, relations, and attributes characterize complementary aspects of entity identity.
- Seed entity alignments are costly and may be unavailable, while prior methods often assume seed relation and attribute alignments can be obtained easily.
- MultiKE divides KG features into complementary views and jointly optimizes view-specific entity embeddings to improve alignment performance.
- MultiKE defines name, relation, and attribute views, applies an appropriate embedding model to each, and combines the resulting representations through three strategies.
- MultiKE introduces cross-KG identity inference at the entity level and at the relation and attribute levels to preserve and enhance alignment between KGs.
- On two real-world datasets, MultiKE largely outperforms existing embedding-based methods, while its views, cross-KG inference, and combination strategies each contribute to improvement.
2 Related Work
Prior KG embedding and entity alignment methods represent graph structure and sometimes refine it with additional information, but they do not treat or extend multiple feature types uniformly. Multi-view representation learning motivates a broader feature-integration approach.
- KG embedding models include translational approaches that interpret relations as vectors translating head entities to tail entities, alongside semantic matching models based on similarity functions.
- Embedding-based entity alignment methods compare entity embeddings across KGs using mappings, unified spaces, self-training, or neighborhood modeling.
- Existing methods use additional features mainly to refine relation-based embeddings and cannot incorporate new features.
- Multi-view representation learning typically identifies multiple views, represents them, and combines their information to learn stronger general representations.
3 Multi-view KG Embedding
The framework learns entity embeddings from complementary name, relation, and attribute views, then uses these representations for entity alignment. It combines view-specific embeddings while modeling literal semantics, relational plausibility, and attribute-value features.
- View Definition: The paper formalizes KGs with three entity views: names, relations, and attributes, represented by N, X, and Y.These views encode entity names, relational links, and attribute-literal pairs, respectively.
- Literal Embedding: An autoencoder compresses concatenated token embeddings into one d-dimensional literal representation.Literals are truncated to five tokens or padded with placeholders before encoding.
- Literal Embedding: Literal embeddings map tokens to pretrained word embeddings when available and character embeddings otherwise.Character embeddings are averaged from embeddings pretrained on the union of KG literal sets.
- Name View Embedding: The name view derives entity embeddings from their extracted names and the learned literal embeddings.The resulting KG embeddings are denoted Θ(1).
- Relation View Embedding: The relation view adopts TransE and scores relation facts by the plausibility of translating head entities to tail entities.Real and corrupted relation facts support logistic training of relation-view embeddings Θ(2).
- Attribute View Embedding: The attribute view uses a CNN over paired attribute and value embeddings to obtain compressed features and trains on attribute-fact plausibility.Negative sampling is omitted because it did not noticeably improve entity alignment.
4 Cross-KG Training for Entity Alignment
MultiKE strengthens entity alignment through cross-KG inference at entity, relation, and attribute levels. It automatically discovers soft relation and attribute alignments during training instead of assuming they are available beforehand.
- Cross-KG inference: MultiKE uses cross-KG entity, relation, and attribute identity inference to capture and enhance alignment information between two KGs.Entity inference relies on seed entity alignment, while relation and attribute inference uses automatically found alignments.
- Entity identity inference: Swapping a seed-aligned head or tail entity in a relation fact creates auxiliary identity-inference probabilities for the corresponding cross-KG fact.These probabilities are maximized over relation facts whose head or tail entities participate in seed entity alignment.
- Attribute identity inference: Attribute-view cross-KG identity inference is defined analogously, using attribute facts whose head entities occur in the seed entity alignment.The attribute inference loss is denoted LCE(Θ(3)).
- Soft relation and attribute alignment: MultiKE automatically finds soft relation and attribute alignments during training rather than assuming strictly equivalent seed alignments.Soft alignment permits similar but non-equivalent schema elements, addressing heterogeneity between KG schemata.
- Soft relation and attribute alignment: Relation similarity combines name similarity from literal embeddings with semantic similarity from relation embeddings using weighted cosine similarities.The weights α1 and α2 are positive and sum to one, with candidate pairs retained when similarity reaches threshold η.
- Training refinement: Soft similarity coefficients reduce the negative effect of inaccurate alignments, while the soft relation alignment is updated iteratively during training.The same cross-KG identity-inference formulation is applied to attributes.
5 View Combination
MultiKE combines view-specific entity embeddings through weighted averaging, shared-space learning, or joint in-training combination. Its training process alternates embedding optimization, soft-alignment updates, and nearest-neighbor entity alignment.
- Combination strategies: MultiKE presents three strategies for combining view-specific embeddings into general entity embeddings.The strategies are weighted view averaging, shared-space learning, and in-training combination.
- Weighted View Averaging: Weighted view averaging assigns each view a weight, lowering the weight of embeddings far from the entity’s average multi-view embedding.This is a late-combination strategy because embeddings are aggregated after independent learning.
- Shared Space Learning: Shared-space learning maps each view-specific embedding space into a common space with an orthogonality constraint on the mapping matrices.Orthogonality preserves distances during transformation, allowing the shared space to learn alignment information from multiple embedding spaces.
- In-Training Combination: In-training combination jointly trains multi-view embeddings so that the views can benefit from one another through agreement with combined embeddings.Unlike the other strategies, this combination participates directly in joint training.
- Combined training process: The complete training process trains literal embeddings first, then learns and combines other views while performing cross-KG identity inference.This sequence integrates view learning, combination, and cross-KG inference within MultiKE training.
- Combined training process: Algorithm 1 initializes name embeddings, repeatedly optimizes the selected combination, updates soft relation and attribute alignments, and finds entity alignments by nearest-neighbor search.The complete process uses pretrained word and character embeddings, Xavier initialization, and AdaGrad optimization.
6 Experiments
Experiments evaluate MultiKE and its variants on two real-world datasets, comparing embedding-based methods, view-specific performance, cross-KG inference, unsupervised alignment, and LogMap.
- Experimental setup: The experiments use DBP-WD and DBP-YG, each containing 100,000 aligned entity pairs with 30% used as seed alignment.
- Experimental setup: MultiKE is evaluated against seven recent embedding-based methods and extended TransD, HolE, and ConvE baselines, using three view-combination variants.The variants are MultiKE-WVA, MultiKE-SSL, and MultiKE-ITC.
- Overall comparison: MultiKE significantly outperforms competing embedding-based entity alignment methods across all metrics on both datasets.On DBP-WD, MultiKE-SSL achieves Hits@1 of 91.86%, a 17.07% absolute gain over BootEA.
- Cross-KG inference: Cross-KG relation and attribute identity inference improves entity alignment, especially for the attribute view.The authors attribute the stronger attribute-view effect to weaker attribute heterogeneity and more accurate soft attribute alignment.
- Unsupervised alignment: MultiKE-ITC achieves acceptable unsupervised alignment results, while the name view does not require seed alignment as supervision.The relation and attribute views can still benefit from seed alignment during training.
- Comparison with conventional methods: MultiKE achieves better results than the competitive conventional method LogMap, demonstrating effectiveness and practicability.
7 Conclusion and Future Work
The paper concludes that multi-view KG embedding with cross-KG inference is effective for entity alignment. It identifies additional views and cross-lingual alignment as future directions.
- Conclusion: MultiKE learns entity embeddings from three representative KG views and introduces two cross-KG training methods for alignment inference.
- Conclusion: Experiments on two real-world datasets demonstrate the effectiveness of the proposed framework.
- Future work: Future work will investigate additional views such as entity types and study cross-lingual entity alignment.