Inferring depictions in natural-language captions for efficient access to picture data

摘要

Multimedia data can require significant examination time to find desired features (“content analysis”). An alternative is using natural-language captions to describe the data, and matching captions to English queries. But it is hard to include everything in the caption of a complicated datum, so significant content analysis may still seem required. We discuss linguistic clues in captions, both syntactic and semantic, that can simplify or eliminate content analysis. We introduce the notion of concept depiction and rules for depiction inference. Our approach is implemented in an expert system which demonstrated significant increases in recall in experiments.