TAT-DQA

Categories: Research, Data Analysis | Pricing: Free | Official Website ↗

TAT-DQA is a large-scale Document Visual Question Answering (VQA) dataset for research on complex, visually-rich documents.

TAT-DQA is a large-scale Document Visual Question Answering (VQA) dataset, an extension of TAT-QA, designed to advance research in question answering over complex, visually-rich documents. It focuses on documents containing both tabular and textual content, particularly those requiring numerical reasoning. The dataset's documents are sourced from real-world financial reports, averaging around 550 words, which is significantly larger than other Document VQA datasets. The dataset includes diverse answer forms such as single span, multiple spans, and free-form, and necessitates various numerical reasoning capabilities like addition, subtraction, multiplication, division, counting, comparison, and sorting. In total, TAT-DQA comprises 16,558 questions linked to 2,758 documents (3,067 document pages). The data is provided in three parts: original PDF documents, converted content from PDFs in JSON format, and QA pairs in JSON format, facilitating research and model evaluation.

Key Features

Large-scale Document VQA dataset
Documents from real-world financial reports
Contains both tabular and textual data
Requires diverse numerical reasoning (addition, subtraction, multiplication, division, counting, comparison, sorting)
Diverse answer forms (single span, multiple spans, free-form)
Provides PDF documents, JSON converted content, and JSON QA pairs
Evaluation script available for model assessment
Leaderboard for model performance tracking

Pros

Large-scale dataset for robust model training and evaluation
Uses real-world financial documents for practical relevance
Combines tabular and textual data, reflecting complex document structures
Requires advanced numerical reasoning, pushing research boundaries
Openly available under a permissive Creative Commons license

Cons

Primarily a dataset, not an end-user tool
Requires technical expertise to utilize for research
Focuses specifically on financial reports, which might limit generalizability to other domains
No direct API for integration into applications
Evaluation and submission process is manual via email

Use Cases

Training and evaluating Document VQA models
Developing algorithms for numerical reasoning over documents
Benchmarking performance of AI models on complex document understanding
Researching multimodal AI for text and table comprehension
Advancing question answering systems for financial analysis

Best For

AI researchers
Machine learning engineers
Academics in NLP and computer vision
Developers working on document understanding systems
Teams focused on financial document processing

Platforms: Web

Watch demo on YouTube ↗

View full TAT-DQA profile on Tools-Radar | Browse Research tools | Alternatives to TAT-DQA

Tools-Radar is a free directory of 10,000+ AI tools — discover, compare, and choose the right AI software for your needs. Visit tools-radar.com