DELICATE: Diachronic Entity LInking using Classes And Temporal Evidence

Technical Summary: DELICATE – Diachronic Entity LInking using Classes And Temporal Evidence

Problem Statement
Entity Linking (EL) in the field of Digital Humanities (DH) and historical text processing faces significant challenges that general-purpose State-of-the-Art (SoTA) models fail to address adequately. These challenges include linguistic variations over time, the necessity of accurate chronological context, and the prevalence of "long-tail" entities under-represented in Knowledge Bases (KBs). Existing domain-specific methods often rely on rigid, manually crafted rules for temporal and type filtering, which lack generalizability. Furthermore, there is a scarcity of benchmark datasets in Italian for training historical EL models, and current approaches utilizing Large Language Models (LLMs) often suffer from high computational costs, hallucination risks, and a lack of interpretability regarding their decision-making processes.

Methodology
The paper introduces DELICATE, a novel supervised architecture designed to perform Entity Disambiguation (ED) on historical Italian texts. The system operates through a two-stage pipeline:

1. Candidate Retrieval (Bi-Encoder): The system utilizes a frozen BERT-based bi-encoder (specifically the BLINKITA model) to encode mention contexts and Wikipedia entity descriptions into a shared latent space. A k-Nearest Neighbors (k-NN) search is performed against a dense vector index (FAISS) containing embeddings of approximately 1.5 million Italian Wikipedia entities to retrieve a top-k candidate list.
2. Candidate Re-ranking (Gradient Boosted Trees): Unlike standard approaches that rely solely on embedding similarity, DELICATE employs a supervised Gradient Boosted Trees (GBTs) classifier to re-rank the retrieved candidates. This classifier ingests a set of pairwise similarity features derived from both the retrieval step and external KBs (Wikidata):
   • Vector Similarity: L2 distances between mention and entity embeddings.
   • String Similarity: Levenshtein and Jaccard distances between surface forms.
   • Type Similarity: A binary match between the mention's entity type (Person, Location, Organization, Work) and the candidate's Wikidata class.
   • Temporal Interval: The time delta between the document's publication date and the entity's earliest associated date in Wikidata (e.g., birth date, inception).

The GBTs model is trained to predict the probability of a correct match, allowing for dynamic assessment of contextual relevance. The system also supports a variant, DELICATE +LLM, where an instruction-tuned open-source LLM (Mistral-Small-24B) is prompted to select the final candidate from the top-10 re-ranked entities, effectively replacing the fixed threshold for NIL (No In Link) prediction.

Key Contributions

1. The DELICATE Architecture: A hybrid approach that combines the efficiency of neural retrieval with the explainability and feature sensitivity of tree-based supervised learning. It explicitly incorporates temporal plausibility and entity type consistency from Wikidata as learnable features rather than hard filters.
2. The ENEIDE Corpus: The authors introduce ENEIDE (Extracted Named Entities from Italian Digital Editions), a multi-domain EL corpus for historical Italian. Semi-automatically extracted from two Scholarly Digital Editions (Digital Zibaldone and Aldo Moro Digitale) spanning the 19th and 20th centuries, it includes literary and political texts with annotations linked to Wikidata.
3. Explainability Analysis: The study demonstrates that the GBT-based approach yields confidence scores with higher correlation to correctness compared to purely neural or generative methods, offering greater interpretability for human-in-the-loop verification.

Experimental Results
The authors evaluated DELICATE on the ENEIDE corpus and the out-of-domain MHERCL-ITA dataset (historical music periodicals), comparing it against baselines including BLINKITA, mGENRE, and ChatEL (an LLM-based approach).

• Performance on ENEIDE: DELICATE variants outperformed all baselines in both Entity Disambiguation (ED) and end-to-end EL tasks. Specifically, DELICATE +LLM achieved the highest micro-averaged accuracy (63.05% on Digital Zibaldone and 72.67% on Aldo Moro Digitale). The model showed particular strength in disambiguating person entities by leveraging temporal cues.
• Performance on MHERCL-ITA: On this domain-shifted dataset with a high proportion of NIL entities (28.67%), DELICATE +LLM achieved a micro-averaged accuracy of 70.44%, significantly outperforming the GBT-only variant and generic baselines. This suggests the LLM component effectively handles high-NIL scenarios where fixed thresholds fail.
• Explainability: Permutation tests revealed that vector similarity and string similarity were the most important features, but temporal intervals remained highly relevant (5th most important). Point-biserial correlation tests showed that DELICATE (without LLM) produced confidence scores strongly correlated with correctness (0.79 on Digital Zibaldone), whereas LLM-based methods showed lower correlation, indicating reduced interpretability of their confidence measures.

Significance and Claims
The paper claims that DELICATE offers a computationally efficient and environmentally sustainable alternative to heavy LLM-based EL systems, particularly for Cultural Heritage institutions with limited resources. By freezing the bi-encoder and training only a lightweight GBT classifier, the system achieves rapid adaptation with minimal data requirements.

The authors assert that their approach complements rather than competes with LLM-based architectures. The integration of structured KB information (dates and classes) via a supervised re-ranker provides a robust mechanism for handling diachronic variations and long-tail entities. Furthermore, the release of the ENEIDE corpus addresses a critical gap in Italian historical NLP, providing a standardized benchmark for future research. The study concludes that while LLMs can enhance candidate selection, especially for NIL prediction, combining them with supervised, interpretable re-ranking yields the most robust and explainable results for historical entity linking.