Representing Context in FrameNet: A Multidimensional, Multimodal Approach

Area of Science:

• Computational Linguistics
• Cognitive Science
• Artificial Intelligence

Background:

• Frame Semantics emphasizes context for meaning construction.
• Berkeley FrameNet computationally represents some contextual aspects.
• Existing models have limitations in capturing full context.

Purpose of the Study:

• Introduce FrameNet Brasil, an enriched FrameNet model.
• Incorporate qualia relations and multimodal data (pictures, videos).
• Address sentence-level cotext and commonsense knowledge computationally.

Main Methods:

• Developed FrameNet Brasil software infrastructure for database construction and corpus annotation.
• Created guidelines for two multimodal datasets annotated for contextual information.
• Designed experiments for frame-evoking unit identification and domain adaptation in Neural Machine Translation.

Main Results:

• Demonstrated FrameNet Brasil's capability to process multimodal input.
• Showcased effective representation of sentence-level context using frames and qualia.
• Validated the importance of structured contextual information over purely form-based manipulation.

Conclusions:

• FrameNet Brasil offers a more comprehensive approach to contextual representation in Frame Semantics.
• Multimodal data enriches FrameNet's ability to capture diverse contextual information.
• Structured contextual representation is crucial for advanced NLP tasks like machine translation.