Proposal for an Extension of Traditional Named Entities: From Guidelines to Evaluation, an Overview

Abstract

Within the framework of the construction of a fact database, we defined guidelines to extract named entities, using a taxonomy based on an extension of the usual named entities definition. We thus defined new types of entities with broader coverage including substantive-based expressions. These extended named entities are hierarchical (with types and components) and compositional (with recursive type inclusion and metonymy annotation). Human annotators used these guidelines to annotate a 1.3M word broadcast news corpus in French. This article presents the definition and novelty of extended named entity annotation guidelines, the human annotation of a global corpus and of a mini reference corpus, and the evaluation of annotations through the computation of inter-annotator agreements. Finally, we discuss our approach and the computed results, and outline further work.