Tables Decoded: DELTA for Structure, TARQA for Understanding
Jahanvi Rajput, Dhruv Kudale, Saikiran Kasturi, Utkarsh Verma, Ganesh Ramakrishnan
Read on arXiv →
Key claim
DELTA and TARQA enhance table understanding with structured text.
In plain English
Imagine you're building a system that needs to understand and interact with tables in documents, like extracting data from financial reports or answering questions about survey results. Currently, many solutions rely on vision-language models that analyze images of tables, which can struggle with accuracy and scalability, especially in multilingual contexts. This often leads to issues like misinterpreting cell contents or failing to recognize the layout correctly, which is what's called table reconstruction failure. Additionally, these models can be cumbersome to adapt for different languages due to their reliance on visual encoders that are not language-agnostic.
To address these challenges, a new approach leverages structured textual representations instead of images. This method, called DELTA, separates the tasks of recognizing physical and logical structures in tables and uses Optical Character Recognition (OCR) to accurately extract both layout and content. The output is in a compact format known as Optimised Table Structure Language (OTSL), which aligns well with large language models (LLMs). The results show that DELTA achieves competitive scores on table structure recognition benchmarks and performs robustly on non-English tables. Furthermore, by fine-tuning an LLM on OTSL sequences, the authors introduce TARQA, which significantly improves performance on table visual question answering tasks. This shift towards structured text representation means that developers can build more efficient and adaptable systems for handling tables across various languages and formats.
The approach introduces a structured textual representation for table understanding, diverging from traditional vision-language models.
The results are validated against established benchmarks, showing competitive performance.
Deep reliability assessment
The methodology supports improved table understanding through structured textual representations and a doubly decoupled approach, but claims of outperforming all VLMs may be overclaimed without specific comparisons.
Reproducibility
yes, the paper mentions releasing code, models, and benchmarks at the provided GitHub URL.
Key figure
Figure 1 illustrates three paradigms for TabVQA: end-to-end VLM, decoupled TSR + OCR for LLM, and the proposed doubly decoupled approach separating physical and logical TSR.