[Paper Review] MERMAID: Metaphor Generation with Symbolism and Discriminative Decoding

Tuhin Chakrabarty, Xurui Zhang+2 more|arXiv (Cornell University)|Mar 11, 2021

Language, Metaphor, and Cognition45 references4 citations

TL;DR

This paper proposes MERMAID, a metaphor generation model that leverages symbolic meaning and discriminative decoding to transform literal sentences into metaphors. By automatically constructing a 93,498-sample parallel corpus using masked language modeling and commonsense reasoning, and fine-tuning BART with a metaphor discriminator, the model generates high-quality metaphors that outperform baselines in human evaluation (66% preference rate) and enhance human-written poetry in task-based evaluations (68% preference).

ABSTRACT

Generating metaphors is a challenging task as it requires a proper understanding of abstract concepts, making connections between unrelated concepts, and deviating from the literal meaning. In this paper, we aim to generate a metaphoric sentence given a literal expression by replacing relevant verbs. Based on a theoretically-grounded connection between metaphors and symbols, we propose a method to automatically construct a parallel corpus by transforming a large number of metaphorical sentences from the Gutenberg Poetry corpus (Jacobs, 2018) to their literal counterpart using recent advances in masked language modeling coupled with commonsense inference. For the generation task, we incorporate a metaphor discriminator to guide the decoding of a sequence to sequence model fine-tuned on our parallel data to generate high-quality metaphors. Human evaluation on an independent test set of literal statements shows that our best model generates metaphors better than three well-crafted baselines 66% of the time on average. A task-based evaluation shows that human-written poems enhanced with metaphors proposed by our model are preferred 68% of the time compared to poems without metaphors.

Motivation & Objective

• To address the lack of parallel training data for literal-to-metaphorical sentence generation.
• To maintain semantic consistency between literal and metaphorical sentences while enabling lexical variation.
• To overcome the tendency of autoregressive models to generate literal over metaphorical language.
• To improve metaphor quality through discriminative decoding guided by a metaphor discriminator.
• To evaluate the model's impact on real-world creative writing tasks, such as enhancing human poetry.

Proposed method

• Construct a parallel corpus of 93,498 [literal sentence, metaphorical sentence] pairs by transforming metaphorical sentences from the Gutenberg Poetry Corpus using masked language modeling and COMET’s commonsense knowledge.
• Use the SymbolOf relation in ConceptNet to ensure symbolic meaning is preserved during literalization, grounding metaphors in shared conceptual symbols.
• Fine-tune the BART sequence-to-sequence model on the constructed parallel corpus to learn literal-to-metaphorical sentence transformation.
• Integrate a discriminative metaphor detector as a reward model during decoding to guide the generator toward more metaphorically appropriate outputs.
• Apply re-ranking with a metaphor detection model (Liu et al., 2019) to improve the quality of generated hypotheses during inference.
• Use human evaluation and task-based assessments to validate the model’s performance on both standalone generation and real-world creative writing.

Experimental results

Research questions

• RQ1 Can symbolic meaning be effectively leveraged to create a large-scale parallel corpus for literal-to-metaphorical sentence generation?
• RQ2 Can a discriminative metaphor discriminator improve the quality and metaphorical fidelity of generated sentences during decoding?
• RQ3 Does the proposed method generate metaphors that are preferred over strong baselines in human evaluation?
• RQ4 Can the model enhance the quality of human-written poetry when used as a creative writing aid?
• RQ5 To what extent does preserving symbolic meaning improve semantic consistency between literal and metaphorical outputs?

Key findings

• The proposed method generated a parallel corpus of 93,498 [literal, metaphorical] sentence pairs using masked language modeling and commonsense reasoning, grounded in symbolic meaning.
• Human evaluation showed that the best MERMAID model generated metaphors preferred 66% of the time over three well-crafted baselines.
• The model outperformed two literary expert baselines in 21% of comparisons and two strong baselines in 81% of cases.
• In a task-based evaluation, poems enhanced with MERMAID-generated metaphors were preferred 68% of the time over poems without metaphors.
• The integration of a metaphor discriminator during decoding significantly improved the quality and metaphorical appropriateness of generated outputs.
• The model demonstrated strong generalization on unseen literal inputs, maintaining semantic consistency while introducing diverse metaphorical expressions.