[Paper Review] Duplication Detection in Knowledge Graphs: Literature and Tools

Elwin Huaman, Elias Kärle+1 more|arXiv (Cornell University)|Apr 17, 2020

Data Quality and Management27 references4 citations

TL;DR

This paper presents a comprehensive literature review and performance evaluation of duplication detection (DD) tools for knowledge graphs (KGs), identifying Duke as the most effective tool for detecting duplicate entities. It proposes a standardized 13-step DD workflow, improvement suggestions for existing tools, and highlights the need for better support in handling conflicting property values and dynamic data.

ABSTRACT

In recent years, an increasing amount of knowledge graphs (KGs) have been created as a means to store cross-domain knowledge and billion of facts, which are the basis of costumers' applications like search engines. However, KGs inevitably have inconsistencies such as duplicates that might generate conflicting property values. Duplication detection (DD) aims to identify duplicated entities and resolve their conflicting property values effectively and efficiently. In this paper, we perform a literature review on DD methods and tools, and an evaluation of them. Our main contributions are a performance evaluation of DD tools in KGs, improvement suggestions, and a DD workflow to support future development of DD tools, which are based on desirable features detected through this study.

Motivation & Objective

• To identify and evaluate existing duplication detection (DD) tools for knowledge graphs (KGs) to address inconsistencies caused by duplicate entities.
• To analyze the challenges in detecting duplicates due to data heterogeneity, typos, and conflicting property values in KGs.
• To propose a standardized, extensible DD workflow to guide future tool development and improve consistency in entity resolution.
• To provide improvement suggestions for current DD tools based on functional and performance evaluation.
• To address gaps in handling conflicting property values and dynamic data over time in KGs.

Proposed method

• Conducted a systematic literature review to identify state-of-the-art DD tools and methods for knowledge graphs.
• Defined a benchmark dataset and created a gold standard for labeling duplicate and non-duplicate entity pairs to enable evaluation.
• Evaluated tools based on functionality, performance, configurability, and support for similarity metrics and fusion strategies.
• Implemented a 13-step DD workflow covering preprocessing, feature selection, normalization, comparison, configuration, and fusion.
• Used similarity metrics (e.g., string similarity) and comparison methods (e.g., blocking, sorted neighborhood) to assess entity matches.
• Applied genetic algorithms and GUI-based verification to refine configurations and validate results in tools like Duke and Sieve.

Experimental results

Research questions

• RQ1 Which duplication detection tools are most effective and efficient in identifying duplicate entities in knowledge graphs?
• RQ2 How do existing tools handle conflicting property values when duplicates are detected?
• RQ3 What are the key limitations in current DD tools regarding usability, configurability, and support for complex similarity metrics?
• RQ4 How can a standardized, reusable workflow improve the development and evaluation of future DD tools?
• RQ5 What improvements are needed to support dynamic data and temporal changes in entity representations over time?

Key findings

• Duke emerged as the most effective tool for duplicate detection, though it was last updated in 2017 and only supports task 1 (identifying duplicates), not conflict resolution.
• FAGI and Sieve were found to support the fusion of conflicting property values, addressing task 2 of the DD problem.
• More than 15 preprocessing functions, 20 comparators, and multiple fusion strategies were identified across evaluated tools, indicating rich but inconsistent implementation.
• String similarity metrics were the most commonly used, but complex metrics are necessary for handling intricate duplication cases.
• Despite adequate documentation, usability issues were observed in most tools, particularly in configuration and GUI interaction.
• The study revealed a need for better support in handling dynamic data and temporal changes in entity properties, especially in crowdsourced KGs.