National and Kapodistrian University of Athens
Effective and efficient entity resolution: from syntactic similarity to large language models
Abstract
dc:descriptionEntity Resolution (ER) is the task of identifying and linking records that refer to the same real-world entity across different data sources. In many applications, entities include textual attributes or are entirely serialized into a single text field and utilize various Natural Language Processing (NLP) techniques—such as similarity functions or Language Models—to detect possible matches. The recent progress in NLP, and particularly the introduction of Large Language Models (LLMs), has significantly advanced the capabilities of ER systems, shifting the focus from surface-level syntactic similarity to deeper semantic reasoning and generalization across domains. This thesis addresses key challenges in Entity Resolution (ER) by leveraging modern NLP techniques, presenting four main contributions. First, TokenJoin introduces a novel method for set similarity joins using maximum weighted bipartite matching, suggesting lightweight and effective token-based filtering. Second, we conduct an in-depth experimental analysis of Small Pre-trained Language Models (SLMs) across multiple ER tasks, such as Blocking and Matching, comparing their effectiveness and efficiency, and examining aspects like generalization and model complementarity. Third, AvengER leverages Large Language Models (LLMs) on multi-candidate prompts in unsupervised settings via ensemble voting and in supervised settings via fine-tuning. Finally, DistillER presents a method for knowledge distillation with LLMs, transferring knowledge from large teacher models to smaller and efficient student models for scalable and cost-effective ER. All methods are evaluated on real-world datasets and established ER benchmarks to ensure comparability with state-of-the-art approaches and strong baselines. To promote transparency and reproducibility, all code, datasets, and experimental results are made publicly available.
Degree
thesis:*- Grantor dc:publisher
- National and Kapodistrian University of Athens
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Zeakis, Alexandros
- Ζεάκης, Αλέξανδρος
Subjects
dc:subject × 11Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60702