Does accompanying text make images easier to find?
This site documents independent research on Multimodal Entity Indexing (MEI): the idea that an image can be treated as a standalone entity in embedding-based search, and that three accompanying text elements (file name, alt text, caption) affect how easily that image is found.
Current status: the conceptual framework and the experimental protocol are drafts. No empirical results have been published. Any effectiveness claim made elsewhere is not yet supported by data that can be inspected.
Research questions
- Does a coordinated set of three text anchors (TACS) increase the chance that an image appears in image search results?
- Which anchor contributes to the effect, alone or in combination?
- Does a measure of image information density (IPR) predict the outcome?
How the work is done
Hypotheses, design, metrics, and the analysis plan are written and registered before any data are collected. All results, including null or negative ones, will be reported, and data and code will be released openly. Details are on the Protocol page.
Three framework concepts
| Term | Short meaning |
|---|---|
| Autonomous Visual Entity (AVE) | A visual asset treated as a knowledge-bearing entity that can be represented on its own in a multimodal vector space. |
| Information-to-Pixel Ratio (IPR) | A proposed measure of the density of factual information in an image relative to its pixel cost. The formula is being reformulated before it is tested. |
| Tri-Anchoring Context System (TACS) | Contextual alignment through three text anchors: file name, deep alt text, and caption. |
These are theoretical proposals awaiting validation, not findings.