A Neuro-Symbolic Self-Supervised Cross-Modal Transformer with Neural Motion Fields and Cross-Linguistic Representation Disentanglement for Scalable and Generalizable Sign Language Translation and Generative Modelling
AVE Trends in Intelligent Computing Systems · 2026
DOI: 10.64091/atics.2026.000282 ↗Kurz gefasst
Fachartikel in der Zeitschrift AVE Trends in Intelligent Computing Systems, 2026. Verfasst von Edwin Shalom Soji und 2 weiteren Autor:innen.
Abstract (Originalsprache)
The study presents a paradigm shift in Sign Language Translation (SLT) and Generative Modelling to overcome the structural and data-scarcity issues of manual communication. The suggested system is a Neuro-Symbolic Self-Supervised Cross-Modal Transformer incorporating Neural Motion Fields (NMF) to learn high-fidelity gesture representations using coordinates and Cross-Linguistic Representation Disentanglement to isolate universal semantic concepts in syntax to language. The fact that the neural output is based on a symbolic logic layer provides linguistic and spatial consistency to the model. The research uses a filtered sample of 457 entries of the RWTH-PHOENIX-Weather 2014T data and the ASL-Lex 2.0 database, which are diasidic sign sequences and phonological attributes. PyTorch, MediaPipe Holistic to extract landmarks, and Weights and Biases were used to implement it and orchestrate the experiment. Findings indicate that the NMF-based architecture outperforms the traditional CNN-LSTM and pure Transformer baselines, achieving significant improvements in BLEU-4 and ROUGE-L scores. In addition, the generative element produces signatures that are more lifelike, while increasing perceptual smoothness by 15 per cent. This paper confirms the hypothesis that a neuro-symbolic system combined with motion-based self-supervision is a scalable step toward a universal, dialect-sensitive sign language technology.
Zitieren
APA
Soji, E. S., Priscila, S. S., & Praveen, B. M. (2026). A Neuro-Symbolic Self-Supervised Cross-Modal Transformer with Neural Motion Fields and Cross-Linguistic Representation Disentanglement for Scalable and Generalizable Sign Language Translation and Generative Modelling. AVE Trends in Intelligent Computing Systems. https://doi.org/10.64091/atics.2026.000282
Vancouver
Soji ES, Priscila SS, Praveen BM. A Neuro-Symbolic Self-Supervised Cross-Modal Transformer with Neural Motion Fields and Cross-Linguistic Representation Disentanglement for Scalable and Generalizable Sign Language Translation and Generative Modelling. AVE Trends in Intelligent Computing Systems. 2026. doi:10.64091/atics.2026.000282
BibTeX
@article{Soji2026,
author = {Soji, Edwin Shalom and Priscila, S. Silvia and Praveen, B. M.},
title = {A Neuro-Symbolic Self-Supervised Cross-Modal Transformer with Neural Motion Fields and Cross-Linguistic Representation Disentanglement for Scalable and Generalizable Sign Language Translation and Generative Modelling},
journal = {AVE Trends in Intelligent Computing Systems},
year = {2026},
doi = {10.64091/atics.2026.000282}
}
EUVIMED