Improving Entity Disambiguation by Reasoning over a Knowledge Base

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2899 - 2912 July 10-15, 2022 ©2022 Association for Computational Linguistics Improving Entity Disambiguation by Reasoning over a Knowledge Base Tom Ayoola∗ Joseph Fisher∗ Andrea Pierleoni Amazon Alexa AI Cambridge, UK {tayoola, fshjos, apierleo}@amazon.com Abstract Recent work in entity disambiguation (ED) has typically neglected structured knowledge base (KB) facts, and instead relied on a limited sub- set of KB information, such as entity descrip- tions or types. This limits the range of contexts in which entities can be disambiguated. To allow the use of all KB facts, as well as descrip- tions and types, we introduce an ED model which links entities by reasoning over a sym- bolic knowledge base in a fully differentiable fashion. Our model surpasses state-of-the-art baselines on six well-established ED datasets by 1.3 F1 on average. By allowing access to all KB information, our model is less reliant on popularity-based entity priors, and improves performance on the challenging ShadowLink dataset (which emphasises infrequent and am- biguous entities) by 12.7 F1. 1 Introduction Entity disambiguation (ED) is the task of linking mentions of entities in text documents to their cor- responding entities in a knowledge base (KB). Re- cent ED models typically use a small subset of KB information (such as entity types or descriptions) to perform linking. These models have strong per- formance on standard ED datasets, which consist mostly of entities that appear frequently in the train- ing data. However, ED performance deteriorates for less common entities, to the extent that many re- cent models are outperformed by outdated fea- ture engineering-based ED systems on datasets that focus on challenging or rare entities (Prova- torova et al., 2021). This suggests models over-rely on prior probabilities, which are either implicitly learned or provided as features, rather than make effective use of the mention context. One reason for this is that the subset of KB information used by the models is not enough to discriminate between *Tom and Joseph contributed equally to this work. similar entities in all contexts, meaning the model has to fall back on predicting the most popular en- tity. Another explanation for the performance drop is that less common entities are prone to missing or inconsistent KB information (e.g. they may not have a description), which is problematic for mod- els which rely on a single source of information. To illustrate, we find that 21% of the 25% least popular1 entities in Wikidata have neither an En- glish description nor any entity type2, leaving no mechanism for models which rely on these two sources of information alone to disambiguate them (other than their label).3 Over half of these entities have at least one KB fact (e.g. [Cafe Gratitude], [headquarters location], [San Francisco]); so by including KB facts the percentage of the least pop- ular entities with no information aside from a label drops from 21% to 8%. Figure 1: Example of a sentence …

Content truncated to 500 words