Traditional Knowledge Graph Embedding (KGE) models excel at mapping discrete relational structures. If a graph contains the triplet (Tokyo, isCapitalOf, Japan), models like TransE learn to treat the relation isCapitalOf as a geometric translation vector shifting the entity Tokyo close to Japan in a latent space.
However, real-world knowledge bases contain more than just abstract entity-to-entity links. They are filled with literals: numerical values, dates, coordinates, and dimensions. A standard knowledge graph might state that Tokyo has a population of 14,000,000 and coordinates of 35.6762° N, 139.6503° E.
Historically, embedding models have suffered from a structural blindspot regarding these attributes. By treating numerical literals either as discrete entity nodes or ignoring them entirely, models lose the intrinsic mathematical properties—such as scale, order, and distance—that define those values.
The Conundrum of Discrete vs. Continuous Data
Why can't we just treat the number 14,000,000 like any other entity node?
In a standard graph topology, every node is distinct. If the population of Tokyo is 14,000,000 and the population of another city is 14,000,005, a traditional relational model views these as two completely independent, orthogonal tokens. The model has no inherent mechanism to understand that these two numbers are nearly identical, or that 14,000,000 is strictly greater than 5,000,000. As noted by Yi Tay et al. in their work on non-discrete attribute prediction, this omission leaves the latent space sparse regarding quantitative characteristics, reducing the model's capacity to represent real-world entities accurately.
When we strip away numerical context, we lose the ability to answer complex queries or predict missing links that depend on physical constraints. For instance, a person cannot be born before their parents, and a small town cannot host an airport larger than Heathrow. To address this, a growing body of research focuses on fusing continuous numerical attributes directly into the geometric spaces of knowledge graph embeddings.
Jointly Embedding Structure and Attributes
The first step toward solving this problem is to model relational triplets and attribute triplets simultaneously within a unified vector space.
Relational Triplet: [Head Entity] ----(Relation)----> [Tail Entity]
Attribute Triplet: [Head Entity] ----(Attribute)---> [Numerical Literal]
In TransEA, Yanrong Wu et al. proposed a joint optimization framework. The model forces entity embeddings to satisfy both relational translations (e.g., $h + r \approx t$) and linear attribute projections (e.g., $W \cdot h \approx a$, where $a$ is the continuous attribute value). By forcing a single entity vector to satisfy both constraints, the structural representation becomes conditioned on the entity's physical attributes.
This joint representation significantly alters the geometry of the latent space. As Jiacheng Huang et al. demonstrate, incorporating literal data types like dates and numbers provides critical context that stabilizes entity representations, leading to much higher precision in downstream tasks like link prediction and knowledge graph completion. Even when predicting purely relational links, knowing the quantitative scale of an entity helps narrow down plausible candidates.
Preserving Mathematical Semantics
While linear projections help, they do not fully capture mathematical relationships like intervals or relative differences. If a model needs to reason about chronological ordering or spatial proximity, it needs an explicit mechanism to calculate numerical distances.
To capture these nuances, Alberto García-Durán et al. introduced KBlrn, a framework that combines latent embeddings with Radial Basis Functions (RBFs) to model numerical features. Instead of forcing a number into a vector space directly, the model evaluates the difference between numerical attributes of two entities through an RBF kernel:
$$\text{Expert}(h, t) = \exp(-\gamma ||a_h - a_t||^2)$$
This allows the system to learn that certain relations (such as isNeighborOf or alteredVersionOf) are fundamentally governed by numerical proximity or specific mathematical scales.
Taking numerical reasoning a step further, Jiaxin Bai et al. developed the Number Reasoning Network (NRN) to handle complex queries involving mathematical constraints like "greater than" or "twice as much." Instead of mapping a number to a single point, the authors represent numerical values using parameterized probability distributions. This probabilistic framing allows the model to perform continuous logical operations over ranges of numbers, enabling systems to execute complex compositional programs over knowledge bases, as explored in datasets like KQA Pro by Shulin Cao et al..
Fusing Literals into Relations
Most attribute-aware models focus on enriching the entity embeddings. However, a relation itself can be defined by the numerical correlations between the entities it connects. For example, the relation hasHigherCostOfLivingThan inherently implies a subtraction or comparison between the financial attributes of the head and tail entities.
[City A: $4000/mo] ----(hasHigherCostOfLivingThan)----> [City B: $2000/mo]
▲
│ (Numerical Fusion)
[Relation Vector enriched by Δ]
To exploit this insight, Antonis Klironomos et al. developed ReaLitE, a plug-and-play method that aggregates numerical attributes from head and tail entities and fuses them directly into the relation embeddings. By explicitly injecting these numerical differentials into the relation vectors during training, the model captures nuanced mathematical patterns that govern whether a link should exist, proving highly effective in scenarios where entity attributes are strongly correlated.
Real-World Impact: Alignment and Summarization
Unlocking numerical awareness solves major bottlenecks in practical knowledge graph applications, most notably in Cross-Lingual Entity Alignment.
When trying to align two knowledge graphs written in different languages (e.g., English Wikipedia and Chinese Baidu Baike), textual entity names and structural topologies can vary wildly. However, numerical values—such as the birth date of a scientist, the height of a mountain, or the geographic coordinates of a city—remain invariant across languages.
As shown by Zhichun Wang et al. and Zequn Sun et al., treating attributes as language-agnostic anchors allows Graph Convolutional Networks (GCNs) to map disparate language graphs into a single, unified vector space. By leveraging character-level attribute embeddings, as implemented by Bayu Distiawan Trisedya et al., models can bypass the unreliability of machine translation entirely, using literal values to align entities with state-of-the-art accuracy.
Furthermore, this multi-faceted representation changes how we compress and query massive graphs. Emetis Niazmand et al. demonstrated that when summarization algorithms can leverage the semantic information contained within non-entity attributes, they can generate highly efficient semantic summary graphs. This reduces graph sizes by up to 99% while maintaining lossless query execution for complex, attribute-dependent lookups.
By moving beyond purely topological abstractions and embracing the continuous, quantitative nature of literal data, modern knowledge graph embeddings achieve a more holistic and grounded understanding of the world they represent.
Key Foundational Work
Knowledge Graph Embedding with Numeric Attributes of Entities (2018) — contributed 4 concept elements to this direction.
KBlrn: End-to-End Learning of Knowledge Base Representations with Latent, Relational, and Numerical Features (2018) — contributed 2 concept elements to this direction.
Cross-lingual Entity Alignment via Joint Attribute-Preserving Embedding (2017) — contributed 1 concept elements to this direction.