Entity linking with relations

In this thesis, you will use pre-trained language models to work on entity linking with a knowledge base.

Open Master Thesis - Contact supervisor for more details!

The task of Entity Linking (EL) aims to link token spans in the text (called mentions) with the corresponding entities in a knowledge base.  Dense representations of mentions and entities are widely used for the task of Entity Linking. Mentions are usually encoded with a local context only, such as the left/right context and the sentence. On the other hand, entities are typically encoded using the entity title and description in Wikipedia [1, 2, 3]. However, some major challenges are largely not tackled in the literature. (C1) Representation of entity mentions with wider context. The mention is usually encoded using the left and right context; we want to experimentwith the impact of using sentence transformers to represent the mention using the text fragment where it appears. (C2) Representation of Knowledge Base (KB) entities. When dealing with a Knowledge Base (KB), we need torethink the representation of entities using triples and how to generate a signif-icant representation. In fact, current works often consider Wikipedia as a KB and use Wikipedia paragraphs to represent entities [1, 2, 3].

The overall idea of this project is to encode mentions using the immediate context (the sentence it appears in) and, on the other hand, encode entities using triples that exist in the KB. Then, by comparing these two representations, we will perform entity linking.

The major research questions that will be addressed to a different extent in the scope of this work are as follows:

  1. Selecting mention sentence encoding approach.
  2. Selecting entity encoding approach.
  3. Candidate generation, meaning selecting entities from the KB that can match the target entity.


  • Knowledge of deep learning
  • Familiarity with pre-trained language models
  • Advanced python programming skills


[1]  D. Cer, Y. Yang, S.-y. Kong, N. Hua, N. Limtiaco, R. S. John, N. Constant,M. Guajardo-C ́espedes, S. Yuan, C. Tar, et al.  Universal sentence encoder. arXiv preprint arXiv:1803.11175, 2018.

[2]  D.  Gillick,  S.  Kulkarni,  L.  Lansing,  A.  Presta,  J.  Baldridge,  E.  Ie,  andD. Garcia-Olano. Learning dense representations for entity retrieval. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 528–537, 2019.

[3]  L. Logeswaran, M.-W. Chang, K. Lee, K. Toutanova, J. Devlin, and H. Lee. Zero-shot  entity  linking  by  reading  entity  descriptions.   In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3449–3460, Florence, Italy, 2019.


To the top of the page