[Paper Review] It's AI Match: A Two-Step Approach for Schema Matching Using Embeddings

Benjamin Hättasch, Michael Truong-Ngoc+2 more|TUbilio (Technical University of Darmstadt)|Mar 8, 2022

Topic Modeling5 citations

TL;DR

This paper proposes a two-step schema matching approach using neural embeddings to improve semantic correspondence detection between heterogeneous databases. By first matching entire tables and then matching individual attributes using contextual embeddings (e.g., Google USE), the method achieves robust, reliable matching even for non-trivial cases like synonymy, multilingual data, or domain-specific terminology, outperforming traditional syntactic and semantic methods on benchmark datasets.

ABSTRACT

Since data is often stored in different sources, it needs to be integrated to gather a global view that is required in order to create value and derive knowledge from it. A critical step in data integration is schema matching which aims to find semantic correspondences between elements of two schemata. In order to reduce the manual effort involved in schema matching, many solutions for the automatic determination of schema correspondences have already been developed. In this paper, we propose a novel end-to-end approach for schema matching based on neural embeddings. The main idea is to use a two-step approach consisting of a table matching step followed by an attribute matching step. In both steps we use embeddings on different levels either representing the whole table or single attributes. Our results show that our approach is able to determine correspondences in a robust and reliable way and compared to traditional schema matching approaches can find non-trivial correspondences.

Motivation & Objective

• To reduce manual effort in schema matching by automating semantic correspondence detection between heterogeneous databases.
• To address limitations of traditional schema matching methods that fail on non-trivial semantic similarities, such as synonyms, multilingual data, or domain-specific terms.
• To evaluate the effectiveness of neural embeddings—particularly contextualized embeddings like Google USE—for matching at both table and attribute levels.
• To demonstrate that combining structural schema information with instance-level data improves matching accuracy and robustness.
• To explore the feasibility and performance of embedding-based matching as a supplement or alternative to existing syntactic and semantic matchers.

Proposed method

• The approach uses a two-step pipeline: first, table-level matching using embeddings of full table representations (schema names, comments, and structural metadata); second, attribute-level matching using embeddings of individual attributes.
• For attribute matching, embeddings are derived from both attribute names and their instance values, with averaging used to create dense vector representations.
• The method leverages pre-trained contextual embeddings such as Google Universal Sentence Encoder (USE) and BERT, with USE showing superior performance in distinguishing similar from dissimilar attributes.
• Instance-based matching is enhanced by sampling to reduce computational cost while preserving discriminative power for large datasets.
• The framework integrates both structural schema features (e.g., data types, constraints) and textual content (names, comments, instance values) into the embedding space.
• A hybrid matching strategy combines results from multiple embedding-based matchers to improve overall accuracy and robustness.

Experimental results

Research questions

• RQ1 Can a two-step embedding-based approach outperform traditional schema matching methods in detecting non-trivial semantic correspondences?
• RQ2 How effective are pre-trained contextual embeddings like Google USE and BERT in matching database tables and attributes across diverse data sources?
• RQ3 To what extent does sampling instance data improve the performance and efficiency of instance-based attribute matching?
• RQ4 How do different embedding sources (e.g., USE vs. BERT) compare in distinguishing semantically similar from dissimilar attributes?
• RQ5 Can embedding-based matching effectively handle challenges such as multilingual data, synonyms, or domain-specific terminology?

Key findings

• The proposed two-step embedding approach significantly improves the detection of non-trivial semantic correspondences, such as synonyms or multilingual equivalents, compared to traditional methods.
• Google USE embeddings outperformed BERT in distinguishing similar from dissimilar attributes, with a wider similarity range and better separation of non-matching pairs.
• On the Adult dataset, the method correctly identified 'marital-status' and 'relationship' as semantically similar attributes, demonstrating strong semantic understanding.
• The approach achieved high accuracy in both table and attribute matching, with computation time under three minutes for a 10-column dataset with 10,000 instances per column on a V100 GPU.
• Sampling instance data improved the discriminative power of embeddings while reducing computational load, especially for large-scale databases.
• The method is robust across diverse data types and performs well even when attribute names are meaningless or semantically ambiguous.