Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Design Decisions for Dual Encoder-based Entity Disambiguation

Susanna Rücker, Alan Akbik

TL;DR

This work focuses on evaluating key design decisions for Dual Encoder-based ED, such as its loss function, similarity metric, label verbalization format, and negative sampling strategy, and presents the resulting model VerbalizED, a document-level Dual Encoder model that includes contextual label verbalizations and efficient hard negative sampling.

Abstract

Entity disambiguation (ED) is the task of linking mentions in text to corresponding entries in a knowledge base. Dual Encoders address this by embedding mentions and label candidates in a shared embedding space and applying a similarity metric to predict the correct label. In this work, we focus on evaluating key design decisions for Dual Encoder-based ED, such as its loss function, similarity metric, label verbalization format, and negative sampling strategy. We present the resulting model VerbalizED, a document-level Dual Encoder model that includes contextual label verbalizations and efficient hard negative sampling. Additionally, we explore an iterative prediction variant that aims to improve the disambiguation of challenging data points. Comprehensive experiments on AIDA-Yago validate the effectiveness of our approach, offering insights into impactful design choices that result in a new State-of-the-Art system on the ZELDA benchmark.

Evaluating Design Decisions for Dual Encoder-based Entity Disambiguation

TL;DR

This work focuses on evaluating key design decisions for Dual Encoder-based ED, such as its loss function, similarity metric, label verbalization format, and negative sampling strategy, and presents the resulting model VerbalizED, a document-level Dual Encoder model that includes contextual label verbalizations and efficient hard negative sampling.

Abstract

Entity disambiguation (ED) is the task of linking mentions in text to corresponding entries in a knowledge base. Dual Encoders address this by embedding mentions and label candidates in a shared embedding space and applying a similarity metric to predict the correct label. In this work, we focus on evaluating key design decisions for Dual Encoder-based ED, such as its loss function, similarity metric, label verbalization format, and negative sampling strategy. We present the resulting model VerbalizED, a document-level Dual Encoder model that includes contextual label verbalizations and efficient hard negative sampling. Additionally, we explore an iterative prediction variant that aims to improve the disambiguation of challenging data points. Comprehensive experiments on AIDA-Yago validate the effectiveness of our approach, offering insights into impactful design choices that result in a new State-of-the-Art system on the ZELDA benchmark.
Paper Structure (61 sections, 2 figures, 15 tables)

This paper contains 61 sections, 2 figures, 15 tables.

Table of Contents

  1. Introduction
  2. The VerbalizED Architecture
  3. Dual Encoder Basics
  4. Mention Encoder.
  5. Label Encoder.
  6. Similarity Computation.
  7. Sampling Negatives for Training.
  8. Inference.
  9. Design Decisions for the Dual Encoder
  10. Enriching representations.
  11. Training Dynamics.
  12. Evaluating Design Choices
  13. Label Verbalization Formats
  14. Design Choices.
  15. Experimental Analysis.
  16. ...and 46 more sections

Figures (2)

  • Figure 1: Overview of VerbalizED during training: The Mention Encoder produces an embedding for each entity mention in a given text (here "Italy" and "Marcello Cuttitta"). The Label Encoder similarly produces an embedding for each unique target in the General Label Set (spanning entities such as "Italy" and "Italy_national_football_team"), by embedding their respective verbalizations. The purpose of training is to learn an embedding space in which mention embeddings lie close to the embeddings of the correct target verbalization. Training uses a Negative Sampling strategy which leverages embeddings to find hard negatives.
  • Figure 2: Iterative variant: All mentions are initially predicted, the top-N with highest similarity are selected for text insertion. The enriched text is then re-embedded, and the remaining mentions are re-predicted.