Reproduce “Towards Text-Image Interleaved Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xin Zhang, Ziqi Dai, Yongqi Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, Jun Yu and 2 more
This research defines Text-Image Interleaved Retrieval (TIIR), a new task for handling queries and documents where text and images are intricately interwoven. It introduces the wikiHow-TIIR dataset and proposes the Matryoshka Multimodal Embedder (MME), which efficiently compresses visual tokens to effectively retrieve interleaved content, outperforming adapted baseline models.
Key insights
- The sequential order and positioning of interleaved text and images are crucial for understanding meaning and achieving effective retrieval, as shuffling content significantly degrades performance.
- An inverted U-shaped relationship exists between the number of visual tokens used and retrieval performance, indicating an optimal balance for information preservation and noise reduction.
- The Matryoshka Multimodal Embedder (MME) effectively addresses the computational challenge of many visual tokens, enabling efficient processing of long interleaved sequences while improving or maintaining accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.