Improving Zero-Shot Entity Linking Candidate Generation with Ultra-Fine Entity Type Information

Proceedings of the 29th International Conference on Computational Linguistics, pages 2429–2437 October 12–17, 2022. 2429 Improving Zero-Shot Entity Linking Candidate Generation with Ultra-Fine Entity Type Information Xuhui Sui1, Ying Zhang1∗, Kehui Song1, Baohang Zhou1, Guoqing Zhao2, Xin Wei2, Xiaojie Yuan1 1 College of Computer Science, TKLNDST, Nankai University, China 2 Mashang Consumer Finanace Co, Ltd {suixuhui,zhangying,songkehui,zhoubaohang}@dbis.nankai.edu.cn {guoqing.zhao02,xin.wei02}@msxf.com, yuanxj@nankai.edu.cn Abstract Entity linking, which aims at aligning ambigu- ous entity mentions to their referent entities in a knowledge base, plays a key role in multiple natural language processing tasks. Recently, zero-shot entity linking task has become a re- search hotspot, which links mentions to unseen entities to challenge the generalization ability. For this task, the training set and test set are from different domains, and thus entity linking models tend to be overfitting due to the ten- dency of memorizing the properties of entities that appear frequently in the training set. We argue that general ultra-fine-grained type infor- mation can help the linking models to learn contextual commonality and improve their gen- eralization ability to tackle the overfitting prob- lem. However, in the zero-shot entity linking setting, any type information is not available and entities are only identified by textual de- scriptions. Thus, we first extract the ultra-fine type information from the entity textual descrip- tions. Then, we propose a hierarchical multi- task model to improve the high-level zero-shot entity linking candidate generation task by uti- lizing the entity typing task as an auxiliary low- level task, which introduces extracted ultra-fine type information into the candidate generation task. Experimental results demonstrate the ef- fectiveness of utilizing the ultra-fine entity type information and our proposed method achieves state-of-the-art performance. 1 Introduction Entity linking (EL) is the task of assigning entity mentions in a text to corresponding entity records in a reference knowledge base. EL plays a key role in the language understanding pipeline, underlying a variety of downstream applications, such as in- formation extraction (Hoffmann et al., 2011; Ji and Nothman, 2016), semantic search (Blanco et al., 2015) and question answering (Berant et al., 2013; Yih et al., 2015; Welbl et al., 2018). In general, EL ∗Corresponding author. Michael would get fouled on every play and still have to play it and just clear himself for shots instead and would rise to that occasion. Michael Jordan Michael Jeffrey Jordan (born February 17, 1963), also known by his initials MJ, is an American former professional basketball player and … Michael Jackson Michael Joseph Jackson (August 29, 1958 – June 25, 2009) was an American singer, songwriter and dancer. Dubbed the "King of Pop", he is regarded as … Extract ultra-fine entity types person, adult, celebrity, performer, artist, musician, professional, singer person, adult, celebrity, performer, actor, athlete, player, basketball player Figure 1: Examples of entity linking with general ultra- fine-grained entity type information. Different ultra- fine-grained types of the two entities are denoted in red. consists of two phases: candidate generation which generates a set of candidates for each mention from millions of entities, and …

Content truncated to 500 words