Introduction
Knowledge graphs (KGs) such as DBpedia and Freebase serve as critical infrastructure for modern AI applications. However, these graphs are frequently siloed by language, hindering the creation of a unified global knowledge base. Network alignment—the process of identifying equivalent entities across different language-specific KGs (e.g., "London" in English and "Londres" in French)—is essential for bridging these gaps.
The Limitations of Translation-Based Models
Historically, network alignment has relied on translation-based embeddings, such as the TransE family. These models operate on the principle that the vector representation of a "head" entity ($h$) plus a relation ($r$) should equal the "tail" entity ($t$): $h + r \approx t$. While computationally efficient, this linear constraint struggles to represent complex local graph topologies, particularly "triangular structures."
As shown in Figure 4, if an entity has multiple relational paths to another entity (e.g., a person is both "born in" a city and also "lives in" that city), translation-based models may incorrectly force different relations to share the same vector or collapse entity representations. This "Triangle Trap" limits the model's ability to capture the unique structural fingerprint of nodes, ultimately degrading alignment accuracy.
The NTAM Framework
To address these limitations, researchers developed NTAM (Non-translational Alignment for Multi-relational Networks). NTAM moves away from rigid geometric constraints in favor of a flexible probabilistic framework.
1. Probabilistic Embedding
Instead of the $h + r = t$ constraint, NTAM defines the probability of a relationship based on the proximity of vectors in a $d$-dimensional space. This approach preserves specific structural motifs, ensuring that the embedding captures the local connectivity patterns of each node without being constrained by linear arithmetic.
2. Joint Embedding via Anchors
To align two disparate language networks, NTAM utilizes "anchors" (known pairs of identical entities or relations). During training, the model employs a "swap" mechanism: it deliberately replaces the representation of an entity with its known counterpart from the other language. This forces the separate networks to overlap within a shared vector space.
3. Efficient Inference
The model is optimized using Stochastic Gradient Descent (SGD) with negative sampling. By using only a single vector per node (rather than separate source and target vectors), NTAM maintains a low parameter count, which helps prevent overfitting and ensures scalability to large datasets.
Experimental Results
NTAM was evaluated on the WK3l-15k and WK3l-120k datasets, which contain cross-lingual data from DBpedia for English, French, and German.
The results demonstrate a significant performance leap over state-of-the-art translation-based baselines like MTransE and ITransE. On the English-French alignment task (WK3l-15k), NTAM achieved a Hits@1 score of 73.6%, compared to just 55.0% for ITransE. Furthermore, NTAM proved robust even with limited training seeds and avoided the error propagation issues common in iterative alignment models.
Conclusion and Future Work
By relaxing the translational constraint, NTAM provides a more expressive and accurate method for cross-lingual knowledge graph alignment. While the current model focuses purely on graph structure, future iterations aim to integrate semantic information—such as entity descriptions and textual attributes—to further enhance alignment in sparse or ambiguous regions of the network.
Summary updated