Boosting Entity Linking Performance by Leveraging Unlabeled Documents

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1935–1945 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1935 Boosting Entity Linking Performance by Leveraging Unlabeled Documents Phong Le1 and Ivan Titov1,2 1University of Edinburgh 2University of Amsterdam lephong.xyz@gmail.com ititov@inf.ed.ac.uk Abstract Modern entity linking systems rely on large collections of documents specifically anno- tated for the task (e.g., AIDA CoNLL). In contrast, we propose an approach which ex- ploits only naturally occurring information: unlabeled documents and Wikipedia. Our ap- proach consists of two stages. First, we con- struct a high recall list of candidate entities for each mention in an unlabeled document. Sec- ond, we use the candidate lists as weak su- pervision to constrain our document-level en- tity linking model. The model treats entities as latent variables and, when estimated on a collection of unlabelled texts, learns to choose entities relying both on local context of each mention and on coherence with other entities in the document. The resulting approach ri- vals fully-supervised state-of-the-art systems on standard test sets. It also approaches their performance in the very challenging setting: when tested on a test set sampled from the data used to estimate the supervised systems. By comparing to Wikipedia-only training of our model, we demonstrate that modeling un- labeled documents is beneficial. 1 Introduction Named entity linking is the task of linking a mention to the corresponding entity in a knowl- edge base (e.g., Wikipedia). For instance, in Figure 1 we link mention “Trump” to Wikipedia entity Donald Trump. Entity linking enables aggregation of information across multiple men- tions of the same entity which is crucial in many natural language processing applications such as question answering (Hoffmann et al., 2011; Welbl et al., 2018), information extraction (Hoffmann et al., 2011) or multi-document summarization (Nenkova, 2008). While traditionally entity linkers relied mostly on Wikipedia and heuristics (Milne and Witten, Mr. Trump discussed Brexit with Mrs. May . Donald_Trump (*) Donald_Trump_Jr. Melania_Trump Ivanka_Trump Trump_(card_games) Trump_(surname) Trump_(video_gamer) Trump_(magazine) Trump,_Colorado ... Brexit(*) May_(singer) May_(surname) Theresa_May (*) Mary_of_Teck Abby_May Cyril_May Fiona_May May_(film) May,_California ... Figure 1: A sentence with candidate entities for men- tions. The correct entities are marked with (*). We automatically extract likely candidates (red bold) and likely negative examples (non-bold red). These are used to train our weakly-supervised model. 2008; Ratinov et al., 2011a; Cheng and Roth, 2013), the recent generation of methods (Glober- son et al., 2016; Guo and Barbosa, 2016; Yamada et al., 2016; Ganea and Hofmann, 2017; Le and Titov, 2018) approached the task as supervised learning on a collection of documents specifically annotated for the entity linking problem (e.g., re- lying on AIDA CoNLL (Hoffart et al., 2011)). While they substantially outperform the tradi- tional methods, such human-annotated resources are scarce (e.g., available mostly for English) and expensive to create. Moreover, the resulting mod- els end up being domain-specific: their perfor- mance drops substantially when they are used in a new domain.1 We will refer to these systems …

Content truncated to 500 words