Abstract
Metaphor interpretation in NLP has traditionally focused on detection or literal paraphrasing. However, a core function of metaphors is highlighting—the process of bringing specific aspects of a concept to the forefront (e.g., "spending time" highlights time as a resource). This paper, presented at EMNLP 2023, introduces the first computational task for identifying these highlighted aspects. By leveraging multitask contrastive learning, specifically a continual learning approach, the authors demonstrate that jointly modeling source domains and highlighted aspects significantly enhances a model's ability to interpret metaphorical intent.
Introduction
In cognitive linguistics, metaphors map a "source domain" (e.g., money) onto a "target domain" (e.g., time). This mapping performs two simultaneous actions: it highlights certain features of the target (e.g., time is valuable and limited) while hiding others (e.g., time is cyclical). While metaphor detection has seen significant progress, the mechanism of highlighting has been largely ignored. Understanding these implicit aspects is crucial for deep comprehension. This work addresses this gap by proposing a task to identify highlighted aspects and exploring how source domain information can assist in this process.
Methods
The researchers utilize Contrastive Learning to associate metaphorical sentences with their corresponding labels—source domains and highlighted aspects—within a shared embedding space. The architecture uses DeBERTa integrated into a Sentence-BERT framework, optimized via Multiple Negatives Ranking Loss to pull correct labels closer to sentence representations in vector space.
Two multitask learning paradigms are compared:
- Continual Learning: The model is fine-tuned sequentially, learning to predict source domains first and then highlighted aspects (or vice versa), allowing knowledge transfer between tasks.
- Joint Learning: The model is trained on both tasks simultaneously using an alternating "round-robin" approach.
Datasets
The study employs the Gordon et al. (2015) corpus, the only metaphor dataset annotated for both source domains and "schema slots" (highlighted aspects). The dataset consists of 1,429 metaphorical sentences with 138 unique source domains and 78 unique highlighted aspects. Its high diversity and sparsity make it a challenging benchmark for classification.
Key Results
The multitask approach consistently outperforms single-task baselines. Notably, continual learning generally proved superior to joint learning, likely because sequential training allows the model to establish a stable representational foundation before adapting to a second sparse task.
In the most challenging setting (without the "literal metaphor" word as a hint), the continual learning approach improved Top-1 accuracy for highlighted aspects by 1.1 points and Top-5 accuracy by 3.6 points. Source domain prediction also saw a boost of 3.4 points. While providing the literal metaphor word (e.g., "killing" in "tax is killing business") significantly improves performance, the multitask benefit persists.
Key Insights and Discussion
- Semantic Synergy: Source domains and highlighted aspects are deeply intertwined. Knowing a metaphor belongs to the war domain makes identifying aspects like threat or struggle much easier.
- Representational Capture: Contrastive learning excels here because it learns concept proximity. Even when the Top-1 prediction is incorrect, the correct aspect frequently appears in the Top-5, indicating the model successfully captures the general semantic neighborhood.
- Future Directions: Future work should address the "hiding" problem—identifying what metaphors obscure—and move toward "open-world" aspect detection, as the current study is limited to a fixed set of labels from the Gordon et al. corpus.
Summary updated