[Paper Review] Integrating and querying similar tables from PDF documents using deep learning

Rahul Anand, Hye-Young Paik+1 more|arXiv (Cornell University)|Jan 15, 2019

Data Quality and ManagementDecision Sciences25 references3 citations

TL;DR

This paper proposes a deep learning-based pipeline for querying and integrating similar financial tables across PDF documents using table type classification and word embedding-based row similarity search. It outperforms traditional text-match methods by achieving 87.84% average hit rate in similar row retrieval using Google News word embeddings, enabling effective, automated analysis of unstructured PDF data without manual schema definition.

ABSTRACT

Large amount of public data produced by enterprises are in semi-structured PDF form. Tabular data extraction from reports and other published data in PDF format is of interest for various data consolidation purposes such as analysing and aggregating financial reports of a company. Queries into the structured tabular data in PDF format are normally processed in an unstructured manner through means like text-match. This is mainly due to that the binary format of PDF documents is optimized for layout and rendering and do not have great support for automated parsing of data. Moreover, even the same table type in PDF files varies in schema, row or column headers, which makes it difficult for a query plan to cover all relevant tables. This paper proposes a deep learning based method to enable SQL-like query and analysis of financial tables from annual reports in PDF format. This is achieved through table type classification and nearest row search. We demonstrate that using word embedding trained on Google news for header match clearly outperforms the text-match based approach in traditional database. We also introduce a practical system that uses this technology to query and analyse finance tables in PDF documents from various sources.

Motivation & Objective

• To enable automated, similarity-based querying of financial tables across heterogeneous PDF documents without manual schema definition.
• To address the challenge of inconsistent table layouts and schemas in PDFs, which hinder traditional data integration methods.
• To develop a practical system that supports rapid data integration and analysis from unstructured, publicly available PDF reports.
• To demonstrate that deep learning-based word embeddings outperform traditional database string similarity for fuzzy row matching in semi-structured data.

Proposed method

• The system converts PDFs to HTML to preserve table structure and enable parsing using HTML tags.
• A deep learning model classifies tables into types (e.g., income statement, balance sheet) based on header content and layout features.
• Word embeddings trained on Google News are used to encode row and column headers into dense vectors for semantic similarity computation.
• Row similarity is computed using cosine distance between embedded header vectors, enabling fuzzy matching across tables with varying nomenclature.
• The approach compares favorably against PostgreSQL’s trigram-based string similarity via pg_trgm for row matching.
• A prototype web interface allows users to query and browse similar rows across multiple PDFs using this embedding-based method.

Experimental results

Research questions

• RQ1 Can deep learning-based word embeddings improve the accuracy of row similarity matching in financial tables extracted from PDFs compared to traditional text-match methods?
• RQ2 How effective is table type classification in grouping similar tables across diverse PDF documents with varying layouts and terminologies?
• RQ3 To what extent can a unified embedding space enable cross-document querying of similar financial table rows without predefined schemas?
• RQ4 Does the use of pre-trained word embeddings (e.g., Google News) outperform custom-trained embeddings for semantic matching in financial table data?
• RQ5 Can a system built on this pipeline enable practical, interactive querying of unstructured PDF data for data analysts without requiring manual data ingestion?

Key findings

• The Google News word embedding model achieved a 87.84% average hit rate in identifying similar rows across tables, significantly outperforming PostgreSQL’s trigram-based similarity search (75.08% hit rate).
• Custom word embeddings achieved a 83.15% average hit rate, showing that pre-trained embeddings are more effective for semantic matching in this context.
• For individual table samples, PostgreSQL occasionally outperformed word embeddings due to strict lexical matching, particularly when vocabulary was sparse or unique.
• Table type classification achieved high accuracy, with clear clustering of table types in the embedding space, indicating effective semantic separation.
• The system demonstrated that deep learning-based semantic matching enables practical, automated querying of financial data across PDFs without manual schema definition.
• The prototype system enables rapid data integration and analysis from public PDF reports, reducing reliance on manual data extraction or premium data sources.