Scaling up Multimodal Pre-Training for Sign Language Understanding.

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Sign language pre-training (SLP) enhances sign language understanding (SLU) but faces limitations in model generalization and neglecting textual cues.
• Existing methods often use task-specific pre-training on small datasets or focus only on visual information, reducing model representational capacity.

Purpose of the Study:

• To develop a multimodal SLP framework that leverages visual context and vision-language consistency for improved sign language video representation.
• To address data scarcity by curating a large-scale text-labeled sign pose dataset (SL-1.5M).

Main Methods:

• Curated a large-scale text-labeled sign pose dataset (SL-1.5M) from diverse sources.
• Proposed a pre-training framework integrating sign-text contrastive learning and masked pose modeling.
• Concurrently modeled manual and non-manual sign language information for holistic visual content representation.

Main Results:

• The framework effectively captures contextual cues in sign pose sequences and aligns semantic text features.
• Achieved new state-of-the-art performance on diverse SLU tasks, demonstrating superior generalization and effectiveness.
• Validated the framework's ability to enhance the representative capability of sign language videos.

Conclusions:

• The proposed multimodal SLP framework significantly improves sign language understanding by integrating visual and textual information.
• The approach overcomes limitations of existing methods, offering better generalization and representational capacity for sign language models.
• This work sets a new benchmark for sign language understanding through advanced pre-training techniques.