Introduction
While Large Language Models (LLMs) have achieved remarkable proficiency in literal language tasks, they often struggle with the creative and abstract nature of figurative language. Generating metaphors is a high-level cognitive task that requires mapping concepts between unrelated domains. This paper introduces MERMAID (MEtaphor geneRation with syMbolism And dIscriminative Decoding), a framework designed to transform literal sentences into high-quality metaphorical ones.
The authors identify three primary hurdles in automated metaphor generation: the scarcity of high-quality parallel corpora (literal-metaphorical pairs), the difficulty of preserving the original sentence's intent (meaning preservation), and the inherent statistical bias of generative models toward common, literal text.
Methodology
Symbolism-Aware Data Construction
To address data scarcity, the researchers developed a novel data-augmentation strategy to build a parallel corpus of approximately 93,000 pairs. Starting with the Gutenberg Poetry Corpus, they identified metaphorical verbs using a BERT-based classifier. To generate corresponding literal versions, they employed Masked Language Modeling (MLM) to suggest candidate verbs.
Crucially, the framework uses COMET (a commonsense knowledge model) to ensure that the literal and metaphorical verbs share the same "symbolic" underpinnings. For example, if a metaphorical sentence uses a verb that evokes "speed" or "energy," the literal replacement must maintain that same symbolic intent. This process ensures the "soul" of the sentence remains intact during the transformation.
Discriminative Decoding
Even after fine-tuning a BART model on the generated corpus, models often drift back toward literal outputs due to their training objectives. To counter this, MERMAID implements a "discriminator" during the inference phase. A separate RoBERTa model is trained to distinguish between literal and metaphorical text. During decoding, the model re-ranks candidate tokens to favor those identified as highly metaphorical by the discriminator, effectively "nudging" the model toward more creative outputs.
Experimental Results
Human and Task-Based Evaluation
MERMAID's performance was evaluated against several baselines using both human judgment and downstream tasks. In head-to-head human evaluations, metaphors generated by MERMAID were preferred over three existing baselines 66% of the time.
In a poetry enhancement task, judges preferred human-written poems that had been "enhanced" by MERMAID (replacing literal verbs with generated metaphors) 68% of the time over the original versions. This suggests that the model can serve as a powerful tool for creative writing assistance.
Qualitative Analysis
As shown in the examples below, MERMAID successfully transforms literal verbs into evocative metaphors while maintaining the context and meaning of the original sentence.
Conclusion
MERMAID demonstrates that metaphor generation can be significantly improved by anchoring transformations in symbolic meaning and using discriminative decoding to overcome literal bias. While currently focused on verb-based metaphors, the framework provides a robust foundation for future work in broader figurative language generation. The success of the model highlights that LLMs possess latent creative potential that can be unlocked with targeted architectural and data-driven interventions.
Summary updated