entitylinking

Entity linking, also known as named entity disambiguation, is the task of connecting textual mentions of entities to their canonical identities in a knowledge base such as Wikipedia or Wikidata. Given a document, the goal is to determine which real-world entity each mention refers to and to attach a stable identifier. The process handles aliasing, where a single surface form maps to multiple entities, and can yield a NIL result when no suitable entity exists in the target knowledge base.

A typical workflow includes three stages: recognition of entity mentions, generation of candidate target entities for each mention, and disambiguation to select the best target. Features used for ranking include local context around the mention, document-level coherence with other linked entities, prior probabilities from corpus statistics, and surface form or type cues. Approaches range from rule-based systems to statistical and neural models, with many systems performing joint recognition and linking. NIL handling is commonly incorporated to indicate unlinked mentions.

Common evaluation datasets and benchmarks include AIDA-CoNLL and TAC-KBP, with metrics such as precision, recall, and F1, and sometimes accuracy for linking decisions. Challenges in entity linking include lexical ambiguity, rare or evolving entities, multilingual and cross-lingual linking, and incomplete knowledge bases. System effectiveness depends on robust recognition, high-quality candidate generation, and accurate disambiguation, as well as maintaining global consistency across documents.

Applications of entity linking span improving search and question answering, enabling knowledge graph construction and data integration, and enhancing information extraction and summarization. It is an active area of research in natural language processing and data science.