Grounding Text in Knowledge Bases

67 concepts63 papers1 proposal

Natural language is inherently ambiguous. A single word can mean entirely different things depending on the context, and crucial background information is often left unsaid. Traditional natural language processing (NLP) treats text as a sequence of isolated strings. To truly understand text, systems must transition from strings to things.

This is the core intuition behind entity linking (or knowledge base grounding): mapping unstructured text spans to unique, canonical identifiers in a structured knowledge base (KB) such as Wikipedia, Wikidata, or domain-specific ontologies. By anchoring ambiguous surface forms to an explicit conceptual space, we can transform isolated text snippets into interconnected nodes within a global web of knowledge.

The Two-Step Bridge: Candidate Generation and Disambiguation

How do we connect a raw string in a document to a specific entity in a database? The process typically relies on a two-stage pipeline: candidate generation and contextual disambiguation.

First, the system must isolate potential entity mentions and find a list of plausible matches within the target knowledge base. As explained by Guilhem Piat et al. in Enriching Contextualized Representations with Biomedical Ontologies: Extending KnowBert to UMLS, this candidate generation step is essential for filtering down millions of database entries into a manageable subset of domain-specific embeddings.

To build this initial bridge, systems often leverage human-curated metadata, such as hyperlinks and redirect pages. For example, Arnab Dutta et al. in Integrating Open and Closed Information Extraction: Challenges and First Steps demonstrated how Wikipedia anchor text can serve as a probabilistic bridge, mapping noisy web-scale surface strings to clean, schema-based ontologies. Similarly, Marc Bron et al. in Related Entity Finding Based on Co-Occurance showed that utilizing internal link structures and redirect data from curated knowledge bases allows systems to bypass the noise and errors common in automated extraction.

Once the candidates are generated, the system must resolve the ambiguity using the surrounding text. In INSIGHTBUDDY-AI: Medication Extraction and Entity Linking using Pre-Trained Language Models and Ensemble Learning, Pablo Romero et al. show how combining fuzzy string matching with keyword-based search can successfully bridge the gap between unstructured clinical text and highly structured medical terminologies like SNOMED-CT.

Why Grounding Works: The Power of Graph Traversal

Why does mapping a string to a database identifier fundamentally change how a machine processes text? The answer lies in the structure of a knowledge base: it is not merely a flat dictionary, but a dense graph of real-world relationships.

Once a text span is linked to a unique entity, a model can immediately traverse the knowledge base's hierarchy to retrieve external factual data, attributes, and categories that are completely absent from the source text. As Emmanuel Cartier et al. observe in Combining Deep Learning Models and Lexical Linked Data: Some Insights from the Development of a Multilingual News Named Entity Recognition and Linking Dataset, linking text to Wikidata allows for deep entity disambiguation and classification by traversing the knowledge base's hierarchy and properties.

[Unstructured Text] ---> "Apple released the new iPhone..."
                             |
                     (Entity Linking)
                             v
[Knowledge Graph]   ---> [Apple Inc.] --(industry)--> [Technology]
                             |
                         (founder)
                             v
                      [Steve Jobs]

This grounding transforms facts from isolated strings into semantically enriched data points connected to world and commonsense knowledge. In CogIE: An Information Extraction Toolkit for Bridging Texts and CogNet, Zhuoran Jin et al. leverage this mechanism to build knowledge-grounded pipelines that integrate world knowledge directly into downstream applications.

Furthermore, this structural context can be injected back into modern language models. In E-BERT: Efficient-Yet-Effective Entity Embeddings for BERT, Nina Pörner et al. demonstrate that aligning entity embeddings with a language model's native vector space allows the model to absorb factual knowledge without expensive retraining. However, the authors note a critical dependency: if the entity linking component fails to map a text span to its unique identifier, the model remains blind to that external factual context.

Taming Messy and Ungrammatical Text

One of the most powerful aspects of entity linking is its ability to act as an anchor for highly informal, ungrammatical, or noisy text. In traditional NLP pipelines, typos or slang can completely derail syntactic parsers. Entity linking solves this by establishing a semantic anchor early in the pipeline.

This intuition was demonstrated early on by Matthew Michelson et al. in Semantic annotation of unstructured and ungrammatical text. Their system, Phoebus, proved that by first linking a messy text string (like an online auction listing) to a known reference database, the system could use the known attributes of the matched entity to guide and constrain the remaining information extraction task.

On social media platforms like Twitter, where text is short and lacks standard grammar, entity linking bridges the "vocabulary gap." As shown by Anh Dang et al. in Real Time Filtering of Tweets Using Wikipedia Concepts and Google Tri-gram Semantic Relatedness, transforming informal text into Wikipedia concepts allows systems to filter and match content in a concept-oriented space rather than relying on exact keyword matches. For complex social media tasks like argument mining, Mihai Dusmanu et al. in Argument Mining on Twitter: Arguments, Facts and Sources combined named entity recognition with external knowledge base links to accurately identify and attribute the real-world sources behind online claims.

Elevating Downstream Reasoning

When an NLP system can ground its vocabulary in real-world facts, its capacity for high-level reasoning increases dramatically. This utility spans across several domains:

By shifting the paradigm from surface-level string matching to deep conceptual grounding, entity linking provides the semantic foundation required for machines to reason about the world the way humans do.

Key Foundational Work

Semantic annotation of unstructured and ungrammatical text (2005) — contributed 1 concept elements to this direction.
Using Encyclopedic Knowledge for Named entity Disambiguation (2006) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_entity_linking_acc3608f,
  title = {entity linking},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/entity-linking/rdr_acc3608fbae4426aac39a68faf9efe80},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_entity_linking_acc3608f_v1,
  title = {entity linking},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/entity-linking/rdr_acc3608fbae4426aac39a68faf9efe80/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake