Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DELICATE: Diachronic Entity LInking using Classes And Temporal Evidence

Cristian Santini, Sebastian Barzaghi, Paolo Sernani, Emanuele Frontoni, Mehwish Alam

TL;DR

DELICATE addresses historical Italian entity linking by uniting a BERT-based bi-encoder with a gradient-boosted trees re-ranker that leverages Wikidata's temporal and type information. It introduces ENEIDE, a diachronic EL corpus spanning the 19th–20th centuries to train and evaluate the approach. Empirically, DELICATE outperforms baseline EL systems and even large LLMs on historical data while offering interpretable confidence scores through its feature-based re-ranking. The work demonstrates strong cross-domain generalization and paves the way for multilingual extensions and integration into knowledge-extraction pipelines for humanities research.

Abstract

In spite of the remarkable advancements in the field of Natural Language Processing, the task of Entity Linking (EL) remains challenging in the field of humanities due to complex document typologies, lack of domain-specific datasets and models, and long-tail entities, i.e., entities under-represented in Knowledge Bases (KBs). The goal of this paper is to address these issues with two main contributions. The first contribution is DELICATE, a novel neuro-symbolic method for EL on historical Italian which combines a BERT-based encoder with contextual information from Wikidata to select appropriate KB entities using temporal plausibility and entity type consistency. The second contribution is ENEIDE, a multi-domain EL corpus in historical Italian semi-automatically extracted from two annotated editions spanning from the 19th to the 20th century and including literary and political texts. Results show how DELICATE outperforms other EL models in historical Italian even if compared with larger architectures with billions of parameters. Moreover, further analyses reveal how DELICATE confidence scores and features sensitivity provide results which are more explainable and interpretable than purely neural methods.

DELICATE: Diachronic Entity LInking using Classes And Temporal Evidence

TL;DR

DELICATE addresses historical Italian entity linking by uniting a BERT-based bi-encoder with a gradient-boosted trees re-ranker that leverages Wikidata's temporal and type information. It introduces ENEIDE, a diachronic EL corpus spanning the 19th–20th centuries to train and evaluate the approach. Empirically, DELICATE outperforms baseline EL systems and even large LLMs on historical data while offering interpretable confidence scores through its feature-based re-ranking. The work demonstrates strong cross-domain generalization and paves the way for multilingual extensions and integration into knowledge-extraction pipelines for humanities research.

Abstract

In spite of the remarkable advancements in the field of Natural Language Processing, the task of Entity Linking (EL) remains challenging in the field of humanities due to complex document typologies, lack of domain-specific datasets and models, and long-tail entities, i.e., entities under-represented in Knowledge Bases (KBs). The goal of this paper is to address these issues with two main contributions. The first contribution is DELICATE, a novel neuro-symbolic method for EL on historical Italian which combines a BERT-based encoder with contextual information from Wikidata to select appropriate KB entities using temporal plausibility and entity type consistency. The second contribution is ENEIDE, a multi-domain EL corpus in historical Italian semi-automatically extracted from two annotated editions spanning from the 19th to the 20th century and including literary and political texts. Results show how DELICATE outperforms other EL models in historical Italian even if compared with larger architectures with billions of parameters. Moreover, further analyses reveal how DELICATE confidence scores and features sensitivity provide results which are more explainable and interpretable than purely neural methods.

Paper Structure

This paper contains 24 sections, 2 equations, 3 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Historical Named Entity Processing
  4. Neural EL in the Humanities
  5. Historical EL Corpora
  6. DELICATE
  7. Bi-encoder and Dense Vector Index
  8. Lookup Table
  9. Gradient-Boosted Trees Classifier
  10. Historical EL using LLMs
  11. ENEIDE Dataset
  12. Experimental Results
  13. Experimental Setup
  14. Evaluation on ENEIDE
  15. Evaluation on MHERCL-ITA
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Representation of a sample sentence from the ENEIDE dataset containing a named entity and candidate entities from Wikipedia presented in a disambiguation page. The image shows how contextual information provided in external KBs allows to better determine the correct entity referenced in a historical text.
  • Figure 2: High-level representation of DELICATE. At first, similar entities with respect to a given mention are retrieved from a dense index of Wikipedia entities by performing k-NN search using the BLINK bi-encoder. In the second step, entities are re-ranked by a GBTs model which takes as input pairwise similarity features computed for each mention-entity pair.
  • Figure 3: Plot of mean importance of each feature of the GBTs model after 30 Permutation Tests on the full ENEIDE test set.