Hedwig: A Named Entity Linker

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4501–4508 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4501 Hedwig: A Named Entity Linker Marcus Klang, Pierre Nugues Lund University Department of Computer Science {marcus.klang, pierre.nugues}@cs.lth.se Abstract Named entity linking is the task of identifying mentions of named things in text, such as “Barack Obama” or “New York”, and linking these mentions to unique identifiers. In this paper, we describe Hedwig, an end-to-end named entity linker, which uses a combination of word and character BILSTM models for mention detection, a Wikidata and Wikipedia-derived knowledge base with global information aggregated over nine language editions, and a PageRank algorithm for entity linking. We evaluated Hedwig on the TAC2017 dataset, consisting of news texts and discussion forums, and we obtained a final score of 59.9% on CEAFmC+, an improvement over our previous generation linker Ugglan, and a trilingual entity link score of 71.9%. Keywords: named entity recognition, named entity linking, named entity annotation 1. Introduction Named entity linking (NEL) is the task of automatically finding and linking mentions of things to unique identifiers. The word thing is too broad for the linkage problem; a more concrete definition used in this paper is linking uniquely separable things, which we can identify by a name, i.e. named entities. The classes of named entities we will then try to link are instances of persons, organizations, locations, etc. Take for instance the named entity of class location: “New York”. This mention can refer to the state1 of New York or the large city2 situated in that particular state. For the latter, a matching unique identifier could be the English Wikipedia label: New_York_City. A typical NEL pipeline consists of many phases includ- ing a name finding, mention detection (MD) phase (e.g. detecting “New York” in a text), a candidate generation (CD) phase (e.g. state or city), and an entity linking (EL) phase (e.g. assigning the label). In addition, these phases might be defined independently (Cucerzan, 2007), or trained jointly (Ganea and Hofmann, 2017). The MD phase is frequently a named entity recognizer (NER), which finds and classifies spans of strings in a set of predefined classes such as persons, organization, location, etc. The CD phase uses the classified mention as input, possibly with context, and from this information generates a list of en- tity candidates. Finally, the entity linking phase ranks and selects the most probable or coherent set of candidate enti- ties. It assigns each mention a label, which corresponds to the unique identifier of the the selected candidate. The unique identifier can be local or global, and its con- crete format is determined by the linker method, which can span the spectrum of fully supervised to unsupervised. This paper uses a supervised approach by linking to prede- termined identifiers. These identifiers are provided by an entity repository which we refer to as the knowledge base (KB). EDL annotated datasets use different kinds of corpora and come with …

Content truncated to 500 words