Entity alignment (EA) is the process of identifying equivalent entities across different knowledge graphs (KGs), such as matching "Apple Inc." in a financial database to "Apple" in a general-purpose KG like DBpedia. Historically, research in this area has operated under a "closed-world" assumption: the belief that every entity in a source graph has a direct counterpart in the target graph.
However, real-world KGs are often constructed from different sources with varying scopes and levels of completeness. As noted by Sun et al. et al. in Knowing the No-match: Entity Alignment with Dangling Cases, this assumption fails in practice, leading to the dangling entity problem. Dangling entities are those that exist in one graph but lack a semantic equivalent in the other. When models are forced to find matches for these nodes, they generate false positives that degrade the reliability of the entire integrated system.
The Failure of the Bijective Assumption
Most early EA models rely on a bijective (one-to-one) mapping constraint. They assume that if entity $A$ from $KG_1$ is the nearest neighbor of entity $B$ in $KG_2$ within an embedding space, they must be the same entity.
This local, greedy approach creates two primary issues:
- Forced Matches: For dangling entities, the model will still pick the "least different" candidate in the target graph. As Ge et al. et al. argue in LargeEA: Aligning Entities for Large-scale Knowledge Graphs, this introduces significant noise, especially when aligning a sparse graph to a dense one where many "distractor" entities may look similar to the source node.
- Many-to-One Conflicts: Without global coordination, multiple source entities often gravitate toward the same popular target entity. Mao et al. et al. demonstrate in From Alignment to Assignment: Frustratingly Simple Unsupervised Entity Alignment that reframing alignment as a global assignment problem—rather than independent similarity searches—is necessary to respect the logical uniqueness of entities.
Detecting the No-Match Case
To handle dangling entities, researchers have shifted toward "dangling-aware" frameworks. Instead of just calculating similarity, these models must learn to abstain from matching.
Sun et al. et al. proposed a multi-task framework that simultaneously learns to align matchable entities and classify others as dangling. This transition from a closed-world to an "open-world" setting is critical for robustness. In Towards Entity Alignment in the Open World: An Unsupervised Approach, Zeng et al. et al. show that unsupervised models can actually outperform supervised ones in these settings by using confidence thresholds to filter out unmatchable nodes that supervised models might be "trained" to force-match.
Active learning has also been adapted to this problem. Liu et al. et al. introduced the concept of "bachelor" entities in ActiveEA: Active Learning for Neural Entity Alignment, noting that standard uncertainty sampling often wastes human effort by asking annotators to label unmatchable entities. By explicitly identifying these bachelors, the system can focus resources on entities that actually have counterparts.
Global Consistency via Optimal Transport
A powerful mathematical tool for solving the dangling entity problem is Optimal Transport (OT). While traditional EA treats matching as a ranking task, OT treats it as a distribution alignment task.
In An Accurate Unsupervised Method for Joint Entity Alignment and Dangling Entity Detection, Luo and Yu utilize a modified Sinkhorn algorithm to enforce one-to-one constraints across the entire graph. Similarly, Tang et al. et al. in A Fused Gromov-Wasserstein Framework for Unsupervised Knowledge Graph Entity Alignment suggest that because real-world graphs are asymmetric, we should use unbalanced optimal transport. This allows the model to "leak" mass—effectively acknowledging that some entities have no destination in the target graph—rather than forcing a uniform distribution where every node must be accounted for.
Impact on Embedding Spaces
Dangling entities do not just cause individual errors; they distort the underlying representation of the graph. Liu et al. et al. point out in Dangling-Aware Entity Alignment with Mixed High-Order Proximities that when a model tries to align a dangling entity, it pulls the embeddings of that entity's neighbors toward incorrect regions of the target graph's vector space.
By using high-order proximities and "matchable priors"—as explored by Yin et al. et al. in Lambda: Learning Matchable Prior For Entity Alignment with Unlabeled Dangling Cases—modern frameworks can pre-identify which nodes are likely to be matchable. This ensures that the embedding process is driven by true correspondences, preserving the structural integrity of the knowledge graphs during the alignment process.
Key Foundational Work
Kernelized Sorting for Natural Language Processing (2010) — contributed 1 concept elements to this direction.
Unsupervised Cluster Matching via Probabilistic Latent Variable Models (2013) — contributed 1 concept elements to this direction.