[Paper Review] Learning Semantic Annotations for Tabular Data

Jiaoyan Chen, Ernesto Jiménez-Ruiz+2 more|arXiv (Cornell University)|May 30, 2019

Data Quality and Management20 references9 citations

TL;DR

This paper proposes a hybrid neural network (HNN) combined with knowledge base (KB) lookup to predict semantic types of entity columns in tabular data without metadata. By leveraging contextual semantics via Att-BiRNN and CNN for intra-column/row features, and P2Vec for inter-column relations from KB reasoning, the method achieves state-of-the-art performance on three web table sets, with strong generalization across datasets despite challenges in cross-domain transfer.

ABSTRACT

The usefulness of tabular data such as web tables critically depends on understanding their semantics. This study focuses on column type prediction for tables without any meta data. Unlike traditional lexical matching-based methods, we propose a deep prediction model that can fully exploit a table's contextual semantics, including table locality features learned by a Hybrid Neural Network (HNN), and inter-column semantics features learned by a knowledge base (KB) lookup and query answering algorithm.It exhibits good performance not only on individual table sets, but also when transferring from one table set to another.

Motivation & Objective

• Address the challenge of predicting semantic types for entity columns in tables lacking metadata such as column names.
• Overcome limitations of lexical matching methods that fail to capture contextual semantics in table content.
• Improve performance by modeling both intra-column and inter-column semantic relationships using neural networks and knowledge base reasoning.
• Evaluate model generalization across different table sets to assess transferability in real-world applications.
• Develop a robust, end-to-end framework for semantic annotation of tabular data that integrates contextual feature learning and KB-based relation inference.

Proposed method

• Use an Attentive Bidirectional Recurrent Neural Network (Att-BiRNN) to embed individual cell phrases within a column, capturing local contextual semantics.
• Apply Convolutional Neural Networks (CNN) over the target column to learn column-level features (intra-column correlations), and over the main cell’s row to learn row-level features (intra-row correlations).
• Introduce a novel KB lookup and query answering algorithm to extract P2Vec—property vectors indicating potential relations between the target column and surrounding columns.
• Combine HNN-learned features with P2Vec to enrich feature representation, enhancing predictive power for column type classification.
• Aggregate predictions from multiple micro-tables (sliding windows over the target column) by averaging their output scores to produce the final prediction.
• Train the model end-to-end using labeled micro-tables extracted from entity columns, with a fixed set of candidate classes for type prediction.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively predict semantic types of entity columns in tables without any metadata, such as column names?
• RQ2 To what extent do contextual semantics from HNN and inter-column relations from KB reasoning improve column type prediction beyond lexical matching?
• RQ3 How well does the proposed model generalize when trained on one table set and tested on a different, unrelated table set?
• RQ4 What is the contribution of individual components—Att-BiRNN, CNN, and P2Vec—to the overall prediction performance?
• RQ5 How do ensemble strategies combining HNN and P2Vec affect robustness and accuracy across diverse table structures?

Key findings

• The proposed method achieves higher overall accuracy than state-of-the-art methods on all three evaluated table sets: T2Dv2, Limaye, and Efthymiou.
• The Att-BiRNN and CNN components of the HNN significantly improve performance, with column-level features showing consistent gains and row-level features providing variable improvements depending on table structure.
• The P2Vec feature, derived from KB lookup and reasoning, provides a substantial performance boost, especially compensating for weak row feature learning.
• Ensemble methods combining HNN and P2Vec yield more robust and improved performance across all datasets, demonstrating the complementary nature of the two components.
• While the model performs strongly within-domain, transfer performance from one table set to another remains challenging, indicating a need for further research in permutation-invariant and generalizable table feature learning.