Wikidata based Location Entity Linking

Online news reading has become general among people and suggesting relevant news articles to readers is a non-trivial task. News recommender systems (NRS) are built to provide appropriate stories to readers based on their interest. News articles usually contain mentions of persons, locations and other named entities which are excellent resources for making sense of readers' news interest. However, entity mentions are often ambiguous. It can make readers retrieve stories that are not relevant to them, impacting the performance of NRS. Entity linking (EL) is a task to extract mentions in documents, and then link them to their corresponding entities in a knowledge base (KB). This task is challenging due to name variations, high ambiguity of entity mentions and incompleteness of the KB. Several approaches have been proposed to tackle these challenges. However, current systems do not focus on improving the performance of EL on location entity mentions which are identified as far more informative entities in news article for user interest profiling. The goal of this paper is to present the design of location entity linking algorithms based on Wikidata KB. We propose new approaches to candidate entity generation and candidate entity ranking of the location EL task. We extensively evaluate the performance of our EL algorithms over a manually annotated AIDA-CoNLL testb news corpus. Experimental results show that our location EL method achieves top-1 precision of 95.58% which is much higher than the state-of-the-art results obtained on the same dataset by collective EL methods.