Description

CoobraTAB is an interactive system that lets blind and visually impaired users explore and analyze complex data tables (financial, epidemiological, scientific) through natural-language dialogue. It ingests tables in a range of formats (HTML, scanned image, PDF, CSV) and reconstructs their full semantic structure — hierarchical headers, merged cells, totalization relationships — so that users can ask questions such as "which are the three companies with the highest revenue?" or "compare France and Germany's results across all quarters."

Main features

• Multi-format ingestion — HTML, images, PDF, CSV, with automatic table-structure recognition.
• Multi-turn dialogue — handles references ("and for Q3?") and analytical intents (comparison, filtering, trend computation).
• User-profile adaptation — adjusts to expertise level, type of visual impairment, exploration style, and desired verbosity.
• Screen-reader optimized output — voice output tuned for JAWS, NVDA, VoiceOver, and TalkBack, plus mathematical braille formatting.
• Grounded answers — responses are anchored in the table's actual data through a grounding mechanism, to avoid hallucination.

Status

Prototypes are under development. A multi-phase user evaluation protocol (formative, technical, cross-format comparative) is planned. Open-source repository and online demo links are forthcoming.

Part of ANR COOBRA → Official project page (external) →