Combining Word and Entity Embeddings for Entity Linking

Abstract

The correct identification of the link between an entity mention in a text and a known entity in a large knowledge base is important in information retrieval or information extraction. The general approach for this task is to generate, for a given mention, a set of candidate entities from the base and, in a second step, determine which is the best one. This paper proposes a novel method for the second step which is based on the joint learning of embeddings for the words in the text and the entities in the knowledge base. By learning these embeddings in the same space we arrive at a more conceptually grounded model that can be used for candidate selection based on the surrounding context. The relative improvement of this approach is experimentally validated on a recent benchmark corpus from the TAC-EDL 2015 evaluation campaign.

Using Knowledge Base Semantics in Context-Aware Entity Linking

“…In these equations, IN e i is the set of incoming hyperlinks in the Wikipedia page for e i , and r is a relationship between e i and e j in a broad sense (a hyperlink for the Wikipedia KB or a relation for an RDF KB). For example, WLM is used in [16,23,28], Ref in [1,23], cosine similarity between entities in [24,25,29,42], while a combination of the three is used in [32]. Beyond the entity relatedness measure, collective linking techniques also differ on their linking strategy, either relying on the direct optimization of Eq.…”

“…Beyond the entity relatedness measure, collective linking techniques also differ on their linking strategy, either relying on the direct optimization of Eq. 3 [21,23,24,29,42] or on a mention-entity graph [1,6,13,16,22,25,28,32,41] such as the one illustrated in Fig. 1.…”

“…They therefore compute an approximation of the solution. [29,42] jointly embed words and entities in the same space and train a classifier with local and global scores to select the best solution. [24] reduces the problem to a submatrix search and solves the optimization of Eq.…”

“…Taking into account the context in which a mention-entity pair appears enables a form of collective linking at a low computational cost, replacing global optimization with an ensemble of local classifications. Following many previous studies, e.g., [6,13,29,32,42], we disregard the named entity mention detection stage, and assume that we know for each document the mentions and their type (person, organization, location, etc.). For the candidate generation stage, a standard architecture is chosen as in [34], relying on Wikipedia for practical reasons.…”

“…There is however only loose semantics behind these interconnections, which prevents from fully taking advantage of KBs reasoning capabilities to better exploit the relationship between entities. To skirt this issue, a few approaches use RDF structured KBs, such as BaseKB, DBpedia or YAGO, instead of Wikipedia [17,29,41]. They thus theoretically benefit from the use of the formal RDF data model for knowledge representation, in particular to exploit the precise semantics of the KB entities (e.g., types) and of their interrelationships (e.g., names and cardinalities), while Wikipedia can just inform about the existence of some unspecified relation between two entities through its hyperlink graph.…”