Source-linked AI summary
Exploring and Evaluating Attributes, Values, and Structures for Entity Alignment
Zhiyuan Liu, Yixin Cao, Liangming Pan, Juanzi Li, Zhiyuan Liu, Tat-Seng Chua
TL;DR
Entity alignment methods have underused attribute triples, while name-biased datasets can overestimate performance. The paper proposes AttrGNN, which jointly models relation and attribute information through attributed value encoding and subgraph partitioning, and introduces a hard evaluation setting. AttrGNN improves average Hits@1 by 5.10% in DBP15k over 12 baselines across cross-lingual and monolingual datasets under regular and hard settings.
Problem
Entity alignment needs better use of attribute triples, and name-biased datasets overestimate performance by favoring equivalent entities with similar names.
Method
AttrGNN jointly models relation and attribute triples using an attributed value encoder, partitioned subgraphs, and dynamically learned attribute and value importance.
Results
5.10% average Hits@1 improvement in DBP15k: AttrGNN outperforms 12 baselines under regular and hard settings across cross-lingual and monolingual datasets.
Takeaways & Limitations
The hard setting selects equivalent entity pairs with very different names, while AttrGNN integrates attribute and relation triples with varying importance for evaluation and performance.
Takeaways & Limitations
The Digital channel performs poorly because numerical calculation is difficult to learn with entity-alignment supervision, so the authors leave it for future work.
Abstract
from arXiv · showhide
Entity alignment (EA) aims at building a unified Knowledge Graph (KG) of rich content by linking the equivalent entities from various KGs. GNN-based EA methods present promising performances by modeling the KG structure defined by relation triples. However, attribute triples can also provide crucial alignment signal but have not been well explored yet. In this paper, we propose to utilize an attributed value encoder and partition the KG into subgraphs to model the various types of attribute triples efficiently. Besides, the performances of current EA methods are overestimated because of the name-bias of existing EA datasets. To make an objective evaluation, we propose a hard experimental setting where we select equivalent entity pairs with very different names as the test set. Under both the regular and hard settings, our method achieves significant improvements ($5.10\%$ on average Hits@$1$ in DBP$15$k) over $12$ baselines in cross-lingual and monolingual datasets. Ablation studies on different subgraphs and a case study about attribute types further demonstrate the effectiveness of our method. Source code and data can be found at https://github.com/thunlp/explore-and-evaluate.
1 Introduction
Entity alignment links equivalent entities across complementary knowledge graphs, but existing methods underuse attribute triples and are evaluated on name-biased datasets. AttrGNN jointly models structural and attribute information with differentiated attribute importance, while a hard setting evaluates pairs with very different names.
- Entity alignment integrates complementary knowledge graphs by linking equivalent entities, supporting question answering, recommendation, and information extraction.
- Equivalent entities often share similar attributes and values, complementing the structural assumption that equivalent entities have equivalent neighbors.For example, equivalent entities can share the Area attribute with similar values.
- Attribute Incorporation Challenge: Existing attribute-aware methods separate relation and attribute networks, preventing discriminative attribute signals from propagating to an entity’s neighbors.Attribute incorporation is more effective when relation and attribute triples are modeled together.
- Attribute Incorporation Challenge: Attributes differ in discriminative power: Name should receive more importance than Time Zone because many cities can share the same time zone.
- Dataset Bias Challenge: DBP15k contains a name-matching bias, with 60%−80% of released equivalent-entity seeds alignable by name matching.Interlanguage links rely heavily on translated entity names, producing many easy pairs.
- Proposed Approach: AttrGNN learns relation and attribute triples jointly, dynamically weights attributes and values, and partitions attributes into separate subgraphs with ensemble strategies.The partitions cover Name, literal, digital, and structural knowledge attributes.
- Evaluation: 5.10% average Hits@1 improvement in DBP15k: AttrGNN outperforms 12 baselines under both regular and hard settings across cross-lingual and monolingual datasets.
2 Related Work
Prior entity-alignment research includes embedding-based and GNN-based approaches, with GNNs leveraging graph structure and neighbor propagation. AttrGNN extends this direction by adding values as graph nodes and using an attributed value encoder for attribute-aware aggregation.
- Entity alignment methods are commonly classified as embedding-based or Graph Neural Network-based approaches.
- Embedding-based Methods: Embedding methods model relation triples in KG embedding spaces, while JAPE, KDCoE, AttrE, and MultiKE incorporate attribute information in different ways.AttrE and MultiKE encode values as extra entity embeddings, but attribute diversity and uninformative values limit performance.
- AttrGNN: AttrGNN’s framework partitions KGs into subgraphs and uses multiple GNN channels to learn their representations.
- GNN-based Methods: GNN-based methods achieve promising entity-alignment results by propagating alignment signals to distant neighbors.Prior work extends relation modeling and matches subgraphs.
- AttrGNN: AttrGNN adds values as graph nodes and uses an attributed value encoder to perform attribute-aware value aggregation.
3 Methodology
AttrGNN partitions each KG by attribute type and uses specialized GNN channels to combine attribute-value signals with relational structure. Channel embeddings are aligned across KGs and ensembled for final entity matching.
- Graph Partition: AttrGNN partitions each KG into four subgraphs for name, literal-value, digital-value, and structure information.The subgraphs share relation triples while using mutually exclusive attribute triples.
- Subgraph Encoder: Different GNN channels encode the four subgraphs, combining attributed value encoders with mean aggregators according to each subgraph’s information type.The Name and Structure channels omit attributed value encoding, while Literal and Digital channels use it before neighborhood aggregation.
- Attributed Value Encoder: The attributed value encoder uses attention to selectively aggregate informative attribute and value features into an entity representation.Attribute and value features are obtained from attribute triples, with BERT used for literal and digital values.
- Mean Aggregator: Mean aggregation combines an entity’s hidden state with neighboring entity features defined by relation triples.The layer applies a learnable matrix, mean operation, and nonlinear activation to produce the next hidden state.
- Graph Alignment: Graph alignment trains each channel to reduce distances between seed-equivalent entities and sampled negatives before producing channel embeddings.Negative samples are selected through nearest-entity search in embedding space, and cosine distance with a margin is optimized.
- Channel Ensemble: The model ensembles channel similarity matrices using either standardized average pooling or learned SVM weights.Average pooling assumes equal channel importance, whereas SVM learns channel weights from sampled similarity scores.
4 Experiments
Experiments evaluate AttrGNN across regular and hard entity-alignment settings, datasets, baselines, and ablations. AttrGNN benefits from combining relation and attribute information, while hard name-dissimilar test sets expose remaining challenges and the value of non-name signals.
- Experimental settings: AttrGNN is evaluated against 12 baselines on cross-lingual DBP15k, monolingual DWY100k, and regular and hard settings.The evaluation uses Hits@N and MRR, with higher values indicating better performance.
- Regular setting: 5.10% Hits@1 and 0.056 MRR improvements over the second-best model are achieved on average in DBP15k.The result is reported for AttrGNN relative to the second-best model and supports integrating attribute and relation triples.
- Regular setting: NameBERT’s strong results and name-based comparisons reveal substantial name bias in DBP15k and especially DBP-YG.The paper reports that NameBERT nearly reaches 100% Hits@1 on DBP-YG and that name signals can be diluted when averaged with other attributes.
- Regular setting: AttrGNN’s improvement over baselines is 8.85% on DBPZH-EN, 1.59% on DBPJA-EN, and 4.86% on DBPFR-EN.The paper attributes the differing gains partly to differences among channel performances and name-related signals.
- Hard setting: AttrGNN remains best in the hard setting but degrades by around 6% for Hits@1, showing that practical entity alignment remains challenging.The hard setting selects equivalent entities with very different names; AttrGNN also shows the lowest degradation among models using entity names.
- Ablation study: Ablations show that non-name features, separate attribute channels, and joint relation-attribute modeling are important, while the Digital channel remains poor.Graph partition improves over MixAttrGNN, the Literal and Structure channels approach the Name channel under the hard setting, and numerical learning is left for future work.
5 Conclusion and Future Work
The paper introduces AttrGNN and a hard experimental setting for entity-alignment evaluation, reporting improvements under regular and hard settings. Future work includes replacing BERT with knowledge-enhanced and number-sensitive text representations.
- AttrGNN integrates attribute and relation triples while assigning them varying importance for entity alignment.
- The hard experimental setting provides a practical evaluation framework for entity alignment.
- Experimental results under regular and hard settings show significant improvements, while the hard setting alleviates severe dataset bias.
- Future work will replace BERT with knowledge-enhanced and number-sensitive text representation models.