SignTrace tackles a difficult dictionary problem: finding a sign when a learner remembers the movement but does not know its meaning or formal feature labels. Users describe the motion in ordinary language, and the system searches 6,699 entries in a Chinese sign-language dictionary.
The pipeline enriches dictionary entries with language models, extracts actions from a query, rewrites descriptions into dictionary-like language, searches across seven channels, and reranks the candidates. On 500 movement descriptions derived from the dictionary, the correct result ranked first 94% of the time and appeared in the top nine 97.4% of the time. Reranking raised first-result accuracy from 71.8% to 94%.
Median processing time was 13.37 seconds when six queries ran concurrently, so the current system is useful as a lookup aid rather than an instant conversational interface. It has entered user trials and received positive informal feedback, but the benchmark’s wording came from the same underlying dictionary and may not reflect how independent users naturally describe unfamiliar signs. The central contribution is an accessible reverse search that connects remembered motion to documented vocabulary without requiring specialist notation.