Neural Entity Linking: A Survey of Models Based on Deep Learning

Tracking #: 2875-4089

Authors: 
Özge Sevgili
Artem Shelmanov
Mikhail Arkhipov
Alexander Panchenko
Chris Biemann

Responsible editor: 
Guest Editors DeepL4KGs 2021

Submission type: 
Survey Article
Abstract: 
In this survey, we provide a comprehensive description of recent neural entity linking (EL) systems developed since 2015 as a result of the "deep learning revolution" in NLP. Our goal is to systemize design features of neural entity linking systems and compare their performance to the prominent classic methods on common benchmarks. We distill generic architectural components of a neural EL system, like candidate generation and entity ranking, and summarize prominent methods for each of them. The vast variety of modifications of this general neural entity linking architecture are grouped by several common themes: joint entity recognition and linking, models for global linking, domain-independent techniques including zero-shot and distant supervision methods, and cross-lingual approaches. Since many neural models take advantage of entity and mention/context embeddings to catch semantic meaning of them, we provide an overview of popular embedding techniques. Finally, we briefly discuss applications of entity linking, focusing on the recently emerged use-case of enhancing deep pre-trained masked language models based on the Transformer architecture.
Full PDF Version: 
Tags: 
Reviewed

Decision/Status: 
Minor Revision

Solicited Reviews:
Click to Expand/Collapse
Review #1
Anonymous submitted on 23/Sep/2021
Suggestion:
Minor Revision
Review Comment:

The authors addressed all my previous comments. One problem I see with current table 2 is that footnotes present in the new column "Learning type for disambiguation" are not present at the bottom of the table. Please fix this.

Review #2
By Italo Lopes Oliveira submitted on 10/Oct/2021
Suggestion:
Accept
Review Comment:

The manuscript is a relevant and unique survey for the Entity Linking research area, analyzing and comparing proposals based on Deep Learning. The survey is easy to understand due to the excellent writing and organization of the sections. Lastly, Entity Linking is an essential topic for the Semantic Web because it enhances text documents or web pages with semantic resources accessible by the Semantic Web.

The following suggestions are not required for the manuscript approval, and their implementations are at the authors' discretion.

* Provide which digital libraries, indexes, and journals were used to collect the papers analyzed in the survey. Query string and other helpful information to replicate the collection process of the papers are welcome;

* Some sets used in the knowledge graph definition are not defined or explained;

* Presents the repository link of the proposal that provides it (if there are enough repositories, a table would be an excellent way to present it);

* Highlight entity relatedness in Section 1 or 2, since it is important enough to be evaluated;

* Future Directions are essential contributions for any survey. It should be expanded and presented before Section 5.

Review #3
By Sahar Vahdati submitted on 26/Oct/2021
Suggestion:
Accept
Review Comment:

The authors have precisely addressed my comments and replied to the comments in the new version of the manuscript that was submitted as 'Survey Article’. The sustainability of the terminologies as well as the presentation of the paper have been significantly improved. The readability and clarity was already in a good level but after application of the requested changed, it is also much better.

Review #4
By Mojtaba Nayyeri submitted on 04/Nov/2021
Suggestion:
Minor Revision
Review Comment:

The authors carefully replied to each of my comments and addressed my concerns with additional explanations. Besides, the replies to reviewers 1 and 3 are also helpful.
However, there are a few points, listed as follows:

=== In Table 6, MAP and nDCG metrics have been used for entity-relatedness evaluation. I would suggest adding formulae for computing these metrics.

=== In equation 1, E_k should be replaced by A_k?

Review #5
By Daza Cruz submitted on 23/Nov/2021
Suggestion:
Accept
Review Comment:

The authors have addressed all my concerns, especially on improving the notation, clarifying the kind of training data required for zero-shot methods, and giving more details on how metrics are computed. I believe these changes have made the paper more comprehensive and clear.