The proceedings of the last LREC2026 conference are online. You can find there information about our latest dataset for emotion analysis of sign language.
DGS-Fabeln-1-SE if an extension of DGS-Fabeln-1 with Sentiment Estimation. We used fours LLMs plus a majority voting filter to associate a sentiment label (negative, neutral, positive) to each segment of the DGS-Fabeln-1 parallel corpus, and we relied on MediaPipe to extract body/face motion features from videos. All together, this makes DGS-Fabeln-1-SE a unique 517 segments parallel corpus among text, sentiment, motion features, and video. Machine learning experiments achieve a 63.1% accuracy in predicting sentiment from motion features using the explainable XGBoost algorithm; highlighting the equivalent importance of body as well as facial features in estimating emotions in Sign Language.
Publication (LREC 2026): https://lrec.elra.info/lrec2026-main-748
Get it from Zenodo: https://doi.org/10.5281/zenodo.18879036
Leave a Reply
You must be logged in to post a comment.