SignTrace
A natural language retrieval system for a Chinese sign language dictionary. It supports users in describing sign gestures in everyday language and querying sign language vocabulary through modules for LLM-based dictionary augmentation, multi-channel retrieval, and candidate reranking.
Timeline (last 90 days)
SignTrace is a retrieval system for Chinese sign language dictionaries that enables reverse lookup of corresponding vocabulary and definitions by describing sign language movements in natural language.
SignTrace's core components include five processing stages: LLM-based dictionary augmentation, motion extraction, dictionary-style rewriting, seven-channel retrieval, and candidate re-ranking.
The SignTrace system covers 6,699 dictionary entries.
Seven-channel retrieval conducts parallel retrieval of candidate entries from multiple semantic dimensions (such as handshape, movement trajectory, body location, and bilateral symmetry) to improve recall coverage.
On a dictionary-derived benchmark test set containing 500 motion description queries, SignTrace achieves a Hit@1 of 94.0%.
SignTrace achieves a Hit@9 of 97.4% on the benchmark test.
SignTrace achieves a Mean Reciprocal Rank (MRR) of 0.9540 on the benchmark test.
The re-ranking stage improves SignTrace's Hit@1 from 71.8% to 94.0%.
The median query processing time for SignTrace under six concurrent queries is 13.37 seconds.
Dictionary-style rewriting converts users' colloquial descriptions into forms closer to standardized dictionary expressions in order to narrow the semantic gap during retrieval.
2 more timeline events
All Facts (12)
SignTrace is a retrieval system for Chinese sign language dictionaries that enables reverse lookup of corresponding vocabulary and definitions by describing sign language movements in natural language.
50%UnverifiedSignTrace's core components include five processing stages: LLM-based dictionary augmentation, motion extraction, dictionary-style rewriting, seven-channel retrieval, and candidate re-ranking.
50%UnverifiedThe SignTrace system covers 6,699 dictionary entries.
50%UnverifiedSeven-channel retrieval conducts parallel retrieval of candidate entries from multiple semantic dimensions (such as handshape, movement trajectory, body location, and bilateral symmetry) to improve recall coverage.
50%UnverifiedOn a dictionary-derived benchmark test set containing 500 motion description queries, SignTrace achieves a Hit@1 of 94.0%.
50%UnverifiedSignTrace achieves a Hit@9 of 97.4% on the benchmark test.
50%UnverifiedSignTrace achieves a Mean Reciprocal Rank (MRR) of 0.9540 on the benchmark test.
50%UnverifiedThe re-ranking stage improves SignTrace's Hit@1 from 71.8% to 94.0%.
50%UnverifiedThe median query processing time for SignTrace under six concurrent queries is 13.37 seconds.
50%UnverifiedDictionary-style rewriting converts users' colloquial descriptions into forms closer to standardized dictionary expressions in order to narrow the semantic gap during retrieval.
50%UnverifiedSignTrace has undergone user trials and received positive informal feedback.
50%UnverifiedSignTrace benchmark queries come from dictionary-derived phrasing and involve pre-selection, which may lead to reduced generalization ability when facing spontaneous user descriptions.
50%