Traditional Entity Linking (EL) and Entity Alignment (EA) systems operate under an assumption of informational abundance. They assume that for every real-world entity, a Knowledge Base (KB) contains a rich text summary (such as a Wikipedia article) or a dense web of well-defined attributes (such as birthdates, occupations, or clear taxonomic categories).
In practice, however, real-world datasets are often highly sparse. Enterprise databases, domain-specific ontologies (e.g., biomedical KBs), and personal data repositories frequently contain millions of entities that lack textual descriptions, fine-grained labels, or structured metadata. When an algorithm encounters a mention like "Lumier" or a rare medical term, it cannot rely on traditional text-to-description matching because the target entity's description is brief, noisy, or entirely absent.
To overcome this metadata sparsity, researchers have shifted away from literal text matching toward methods that exploit structural topology, cross-domain knowledge transfer, and minimal contextual anchors.
Synthesizing Context from Graph Topology
When an entity lacks a textual summary, its position within a network graph becomes its primary identifying feature. If we cannot compare text descriptions, we can instead compare the neighborhoods or relational contexts of the entities.
Besançon et al. addressed description-free entity linking by constructing a "relational context"—essentially a pseudo-document built from an entity's graph neighbors. By aggregating the names and attributes of connected entities, the authors created a proxy text representation that standard disambiguation algorithms could process.
Building on this idea, Wagner et al. bypassed text generation entirely by learning distributed representations (embeddings) of the relations themselves. By running random walks over a knowledge graph, the authors treated the graph structure like a text corpus, allowing the model to capture latent semantic similarities between entities based purely on their structural roles rather than their missing textual descriptions.
Traditional EL: [Mention Context] <---> [Entity Text Description] (Fails if empty)
Structural EL: [Mention Context] <---> [Graph Neighborhood / Relation Embeddings]
For cross-lingual entity alignment—where language barriers exacerbate text sparsity—relying solely on literal attributes fails. Yang et al. demonstrated a multi-aspect framework that combines Graph Convolutional Networks (GCNs) with multilingual language models. When literal descriptions are missing, the GCN captures the underlying graph topology to provide the discriminative power necessary to align identical entities across different languages.
Cross-Domain Knowledge Transfer
Another powerful strategy for mitigating local sparsity is to "borrow" structure from a denser, public knowledge base. Domain-specific KBs (such as those in finance or medicine) are notoriously sparse because specialized terms or rare entities rarely receive comprehensive manual curation.
As Varma et al. observed, biomedical KBs often lack the fine-grained categorical labels needed to distinguish between highly similar concepts. To resolve this, the authors introduced a cross-domain integration method that maps rare medical entities to Wikidata. By linking the sparse local entity to a rich general-domain graph, the model inherits a dense web of structural relationships, resulting in massive accuracy gains for rare entity disambiguation.
Similarly, when dealing with automatically harvested web taxonomies that lack clean attribute data, Lee et al. formulated entity resolution as a multi-way graph cut problem. By mapping the noisy, sparse taxonomy to a curated dataset and leveraging both positive and negative web evidence, the authors demonstrated that structural alignment can cleanse and enrich datasets containing billions of sparse entity pairs.
Disambiguation with Minimal Anchors
In the most extreme cases of sparsity, a knowledge base may contain no graph structure at all—only a flat list of names. This scenario frequently arises in disaster response tracking or niche product monitoring.
To resolve mentions against a flat list, Lin et al. proposed using "seed mentions." The intuition is that while many mentions in a document are highly ambiguous, a few can be linked with high confidence based on clear document co-occurrences. These confident "seeds" then act as a bridge, providing the contextual clues necessary to disambiguate the remaining, more difficult mentions against the list.
When brief snippets of text are available, models can maximize their utility by extracting implicit hierarchical signals. Sui et al. showed that zero-shot entity linking can be improved by training models to predict "ultra-fine" entity types (e.g., identifying an entity as a basketball player rather than just a person) from sparse descriptions. This granular typing constrains the candidate search space, preventing the model from making broad alignment errors.
The Extreme Sparsity of Personal Data
The problem of attribute sparsity reaches its peak in Personal Information Management (PIM) systems, which integrate highly fragmented data like emails, calendars, and contact lists. Unlike public KBs, a personal contact entry might consist of nothing more than a single email address or a nickname.
Early foundational work by Levy et al. on the SEMEX platform highlighted that personal records lack global identifiers, making automated association a unique challenge. Because traditional graph-alignment algorithms fail when entities share almost no overlapping properties, specialized matching techniques are required.
Montoya et al. tackled this in their Thymeflow framework by combining string similarity metrics with Inverse Document Frequency (IDF) weights. The intuition is straightforward: if two sparse records share a highly unique string (like an unusual nickname or a specific email domain), that shared feature is weighted heavily enough to justify an alignment, bypassing the need for the dense multi-attribute profiles required by standard corporate database resolution.
Key Foundational Work
Improving Author Coreference by Resource-Bounded Information Gathering from the Web (2007) — contributed 1 concept elements to this direction.
A Platform for Personal Information Management and Integration (2005) — contributed 1 concept elements to this direction.
JHU1 : An Unsupervised Approach to Person Name Disambiguation using Web Snippets (2007) — contributed 1 concept elements to this direction.