Learning contextualized semantics from co-occurring terms via a Siamese architecture.

Area of Science:

• Multimedia Information Retrieval
• Natural Language Processing
• Machine Learning

Background:

• Bridging the semantic gap in multimedia information retrieval is crucial for accurate concept understanding.
• Out-of-vocabulary (OOV) concepts present a significant challenge in modeling semantics within context.

Purpose of the Study:

• To develop a novel method for learning contextualized concept semantics from descriptive terms.
• To address the semantic gap and improve the representation of OOV concepts in multimedia data.

Main Methods:

• A novel Siamese architecture is proposed to model contextualized semantics from descriptive terms.
• Unsupervised learning, pattern aggregation, and probabilistic topic models are employed to capture co-occurring term semantics.
• A concept embedding space is generated for in-context term representation, accommodating OOV concepts.

Main Results:

• The developed Siamese architecture effectively captures contextualized semantics, creating a robust concept embedding space.
• Visualization demonstrates the properties of the learned concept embedding space.
• Experimental evaluation on semantic priming across diverse datasets shows superior performance compared to state-of-the-art methods.

Conclusions:

• The proposed approach successfully models contextualized semantics and addresses the challenge of OOV concepts in multimedia retrieval.
• The learned concept embedding space facilitates better understanding and retrieval of multimedia information.
• The method demonstrates significant improvements in semantic priming tasks across multiple domains.