In the static world of traditional databases, an entity is often treated as a fixed point: a person has a name, a city has a location, and a company has a brand. However, real-world data is rarely static. Over decades or centuries, cities are renamed (St. Petersburg becomes Leningrad and then St. Petersburg again), researchers change affiliations, and organizations merge or split.
This phenomenon, often called temporal drift, creates a fundamental challenge for information systems. If a system cannot recognize that "Ceylon" and "Sri Lanka" refer to the same geographic entity, its knowledge becomes fragmented. Researchers are now developing "diachronic" reasoning techniques—methods that track identity through time—to bridge the gap between historical records and modern knowledge bases.
The Challenge of Evolving Concepts
Most entity linking systems are designed to map a mention in text to a contemporary reference, like a current Wikipedia page. But as Arianna Graciotti and colleagues point out in their work on KE-MHISTO, historical documents often contain "long-tail" entities that no longer exist in modern digital records.
When an entity’s attributes change—such as a person’s name or a company’s rebranding—automated systems struggle to maintain identity consistency. Alon Y. Halevy and Xin Dong 0001 highlighted this in their work on SEMEX, noting that personal information management requires reconciling disparate mentions over long periods without manual intervention. Without this, a single historical subject is treated as multiple unrelated entities, a problem Nina Tahmasebi identifies as a major hurdle for "culturomics"—the study of cultural trends through text Visions and open challenges for a knowledge-based culturomics.
Anchoring Identity with Functional Context
How can a machine know that two different names refer to the same thing across different eras? One powerful intuition is to look at what the entity does rather than what it is called.
Amal Chaminda Kaluarachchi and Aparna S. Varde introduced the concept of SITACs (Semantically Identical Temporally Altering Concepts) in Intelligent Time-Aware Query Translation for Text Sources. They propose using verbs as "functional anchors." While the noun (the entity name) might change, the predicate-argument structures—the actions associated with the entity—often remain stable. By analyzing shared events and relationships, systems can automatically translate modern search queries into their historical equivalents.
Cross-Version Mapping and Redirects
Another strategy involves leveraging the evolution of the knowledge bases themselves. Knowledge bases like Wikipedia are not static; they are living documents with their own histories of redirects and revisions.
Silviu Cucerzan demonstrated that using a contemporary, comprehensive knowledge base can actually help resolve entities in older documents TAC Entity Linking by Performing Full-document Entity Extraction and Disambiguation. By extracting entities using a modern snapshot and then mapping them back to a legacy reference version using title redirects, systems can use "future" knowledge to clarify "past" ambiguities.
This multi-step alignment is crucial because, as Iryna Gurevych and Nicolai Erbs found, the redirect structures and entity definitions in a knowledge base drift over time UKP-UBC Entity Linking at TAC-KBP. Reconciling these temporal snapshots ensures that a mention is linked to the correct historical identifier even if the underlying reference material has been reorganized.
Managing Merges and Splits
Identity is not always a one-to-one mapping over time. Entities can merge (two companies becoming one) or split (a colony becoming multiple independent nations). Barbara Bazzanella and Themis Palpanas argue that a robust entity representation model must account for these transitions while maintaining historical provenance Towards a General Entity Representation Model.
In community-scale systems like DBLife, developed by AnHai Doan and Raghu Ramakrishnan 0001, this is handled by treating the community as a dynamic entity-relationship graph. By combining automated extraction with community-driven "mass collaboration," the system tracks researchers as they move between institutions, ensuring that the "entity" of a researcher remains consistent despite changes in their professional metadata.
By treating time as a first-class dimension rather than a static attribute, these researchers are moving toward a more "fluid" understanding of data—one where identity is a thread that persists even as the labels we use to describe it change.