SlideVQA: Multi-Image Document Question Answering on 10K+ Slide Decks

SlideVQA is a document visual question answering dataset containing between 10,000 and 100,000 records, released by NTT-hil-insight in 2023. It focuses on multi-image reasoning where models must select specific evidence slides from a deck to answer natural language questions.

Use Cases

• Evidence image selection: Identifying specific slide indices that contain the information needed for a query
• Multi-modal answer generation: Producing text answers by synthesizing visual data from multiple slides
• Slide-level retrieval: Ranking images within a deck to find the most relevant visual context

Strengths

• 10,000 to 100,000 record scale
• Multi-image context requiring evidence retrieval
• Peer-reviewed methodology from Arxiv 2301.04883

Limitations

• Domain-specific focus on presentation slides may not generalize to other document types like forms
• High memory requirements for processing multiple images per query

Provenance

Source

NTT-hil-insight (Arxiv 2301.04883)

Collection Method

Annotated slide decks for evidence selection and question answering

Freshness

Last updated March 2025; based on 2023 research.