Re-Representing Metaphor: Modeling Metaphor Perception Using Dynamically Contextual Distributional Semantics.

Area of Science:

• Computational Linguistics
• Cognitive Science
• Natural Language Processing

Background:

• Distributional semantics represent words in high-dimensional spaces based on co-occurrence.
• Static, global word representations struggle to capture context-dependent meaning and metaphor.
• Modeling metaphor requires understanding how word meanings shift in different contexts.

Purpose of the Study:

• To present a novel context-dependent approach for modeling word meaning, specifically applied to metaphor.
• To hypothesize that dynamic, contextualized word representations improve metaphor modeling compared to static approaches.
• To evaluate the effectiveness of the proposed model against state-of-the-art static models.

Main Methods:

• Developed a novel approach that dynamically projects low-dimensional subspaces for contextualized word representations.
• Re-represented words in ad hoc spaces to capture context-specific geometrical and conceptual configurations.
• Tested the model on a dataset of English word dyads rated for metaphoricity, meaningfulness, and familiarity.

Main Results:

• The proposed context-dependent model significantly outperformed a state-of-the-art static model in capturing human ratings of metaphoricity, meaningfulness, and familiarity.
• The model's effectiveness was correlated with the amount of contextualizing work involved in the re-representational process.
• Demonstrated that dynamic re-representation is crucial for modeling the semantics of metaphor.

Conclusions:

• Context-dependent word representations offer a more effective approach to modeling word meaning, particularly for figurative language like metaphor.
• The dynamic projection of subspaces allows for flexible and context-specific semantic modeling.
• This approach advances the field of computational semantics by providing a more nuanced understanding of word meaning in context.