Context-enhanced Adaptive Entity Linking

Context-Enhanced Adaptive Entity Linking Filip Ilievski∓, Giuseppe Rizzo∗, Marieke van Erp∓, Julien Plu◦, Raphaël Troncy◦ ∓Vrije Universiteit Amsterdam, Netherlands {f.ilievski,marieke.van.erp}@vu.nl ∗ISMB, Italy giuseppe.rizzo@ismb.it ◦EURECOM, France {plu,troncy}@eurecom.fr Abstract More and more knowledge bases are publicly available as linked data. Since these knowledge bases contain structured descriptions of real-world entities, they can be exploited by entity linking systems that anchor entity mentions from text to the most relevant resources describing those entities. In this paper, we investigate adaptation of the entity linking task using contextual knowledge. The key intuition is that entity linking can be customized depending on the textual content, as well as on the application that would make use of the extracted information. We present an adaptive approach that relies on contextual knowledge from text to enhance the performance of ADEL, a hybrid linguistic and graph-based entity linking system. We evaluate our approach on a domain-specific corpus consisting of annotated WikiNews articles. Keywords: Adaptive, Contextual, Entity Linking, Knowledge Extraction 1. Introduction The Linked Open Data (LOD) cloud,1 i.e. the union of nu- merous interlinked RDF datasets published following the linked data principles, represents a remarkable source of real-world knowledge. As such, it can be exploited by intelligent systems to extract information from textual re- sources. In particular, an intelligent system that anchors entities from text to existing entities in a knowledge base, namely an entity linking system, could benefit from the structured semantic knowledge describing those entities in the knowledge base. This knowledge enables an entity link- ing system to model and perform reasoning over the seman- tic context of the real-world entities it links to. However, the heterogeneity of the LOD knowledge also represents a barrier towards its automatic use, integration and manipulation. Entity linking systems have thus dif- ficulties to adapt across domains, i.e. to perform uni- fied processing of textual resources from different domains. This has inspired the development of domain-focused sys- tems that perform reasonably well on concrete, domain- dependent information extraction tasks, but demonstrate low accuracy when applied in different domains. The key intuition we follow in this paper is that entity link- ing can be customized for new domains by making use of contextual knowledge associated with the textual content to analyze and the task that will make use of the extracted in- formation. We investigate how to enhance the cross-domain performance of entity linking through contextual adapta- tion, i.e. the integration of semantic knowledge from a domain-specific knowledge base with evidence from addi- tional information sources (the text to analyze and the task to address). Our approach aims to enhance the cross-domain perfor- mance of ADEL, a hybrid linguistic and graph-based en- tity linking module by applying a set of heuristics for con- textual adaptation. These heuristics consider the order of processing of text, coreference of entity mentions, topical domain relevance and semantic typing. We assess the im- 1http://lod-cloud.net pact of our approach on an automotive domain dataset, the MEANTIME corpus (Minard et al., 2016), as well as on the standard AIDA-YAGO2 dataset …

Content truncated to 500 words