Improving Candidate Retrieval with Entity Profile Generation for Wikidata Entity Linking

Findings of the Association for Computational Linguistics: ACL 2022, pages 3696 - 3711 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Improving Candidate Retrieval with Entity Profile Generation for Wikidata Entity Linking Tuan Lai, Heng Ji, ChengXiang Zhai University of Illinois at Urbana-Champaign {tuanml2, hengji, czhai}@illinois.edu Abstract Entity linking (EL) is the task of linking en- tity mentions in a document to referent enti- ties in a knowledge base (KB). Many previ- ous studies focus on Wikipedia-derived KBs. There is little work on EL over Wikidata, even though it is the most extensive crowdsourced KB. The scale of Wikidata can open up many new real-world applications, but its massive number of entities also makes EL challeng- ing. To effectively narrow down the search space, we propose a novel candidate retrieval paradigm based on entity profiling. Wiki- data entities and their textual fields are first indexed into a text search engine (e.g., Elas- ticsearch). During inference, given a men- tion and its context, we use a sequence-to- sequence (seq2seq) model to generate the pro- file of the target entity, which consists of its title and description. We use the profile to query the indexed search engine to retrieve candidate entities. Our approach complements the traditional approach of using a Wikipedia anchor-text dictionary, enabling us to further design a highly effective hybrid method for candidate retrieval. Combined with a sim- ple cross-attention reranker, our complete EL framework achieves state-of-the-art results on three Wikidata-based datasets and strong per- formance on TACKBP-20101. 1 Introduction Entity linking (EL) is the task of mapping entity mentions in a document to standard referent enti- ties in a target knowledge base (KB) (Dill et al., 2003; Cucerzan, 2007; Mihalcea and Csomai, 2007; Milne and Witten, 2008; Ji et al., 2010; Radhakr- ishnan et al., 2018; Lai et al., 2021a; Jiang et al., 2021). EL systems have found applications in many tasks such as question answering (Li et al., 2020a), knowledge base population (Dredze et al., 2010), 1 Our system is publicly available at https://github. com/laituan245/EL-Dockers/. information extraction (Li et al., 2020b; Wen et al., 2021; Lai et al., 2021b), and query interpretation (Kasturia et al., 2022). In general, the task is chal- lenging because the same word or phrase can be used to refer to different entities. At the same time, the same entity can be referred to by different words or phrases. Given the importance of EL, researchers have introduced a plethora of EL methods, ranging from using hand-crafted features (Ratinov et al., 2011; Pan et al., 2015) to using deep language models (Agarwal and Bikel, 2020; Cao et al., 2021; Botha et al., 2020). The vast majority of these studies have focused on linking mentions to Wikipedia or Wikipedia-derived KBs such as DBpedia (Auer et al., 2007) or YAGO (Suchanek et al., 2007). As of November 2021, there are about 6.4 million arti- cles in English Wikipedia. However, many entities are still missing from Wikipedia (Redi et al., 2020). On the other hand, Wikidata, the most …

Content truncated to 500 words