A collective approach to ranking entities for mentions

Entity linking (EL) is the task of mapping name mentions in web text to their entities in a knowledge base. Most of earlier EL work in the knowledge based approach is usually formulated as a ranking problem, either by (i) non-collective approaches with supervised models, or (ii) collective approaches by leveraging global topical coherence which means semantic relations between entities through graph-based approaches. For the mapping process, we can regard it as selecting an entity to its mention by combining these two methods. In this paper, we propose a probabilistic model that ranks related entities to name mentions where ranking is customized by using three types of data: popularity knowledge of the entity, context similarity between mentions and the entity, and semantic relations between mapping entities. Specifically, we first propose an EL model utilizing global topical coherence that means semantic relatedness between entities, as well as using local mention-to-entity compatibility, to improve recall and precision. The key benefit of our model comes from 1) combination of two methods to provide customized ranking for mentions, 2) the model can save a large amount of calculation by efficiently finding candidate combinations of entities through global semantic coherence.