Metaphor Detection with CrossLingual Model Transfer

Metaphors are pervasive in everyday language, literature, and specialized discourse. They enable speakers to convey complex ideas by linking concrete source domains with abstract target domains (e.g., time is money). Automatic detection of metaphorical language is a crucial step for downstream tasks such as sentiment analysis, machine translation, and narrative understanding. While there has been considerable progress for highresource languages like English, most languages lack the annotated resources needed to train reliable detectors.

Why CrossLingual Transfer?

Creating goldstandard metaphor annotations is laborintensive. A single wellannotated corpus for a language can take months of expert work. Crosslingual model transfer offers a way to leverage rich resources in a source language (usually English) to bootstrap detectors for target languages with limited or no data. The central idea is to train a model on a highresource language and adapt it to other languages through multilingual representations.

Core Components of a Transfer Pipeline

1. Multilingual Embeddings: Vector spaces where words from different languages share a common semantic map. Wordlevel embeddings (e.g., MUSE, fastText) and contextual models (e.g., mBERT, XLMR, or the newer BLOOMz) provide the foundation for transfer.
2. Annotation Projection: When parallel corpora with sentencelevel alignments are available, metaphor labels from the source side can be projected onto the target side using word alignment tools.
3. FineTuning Strategies:
   • ZeroShot: Directly apply the sourcetrained model to the target language.
   • FewShot: Finetune on a small set of manually annotated target examples.
   • Adversarial / Contrastive Learning: Reduce languagespecific variance while preserving metaphorspecific features.
4. Evaluation: Standard metrics (precision, recall, F1) on a heldout targetlanguage test set, possibly supplemented by human judgment for ambiguous cases.

Recent Research Highlights

Multilingual BERT for Metaphor Detection (2021) Researchers finetuned mBERT on the VUMetBank (English) and evaluated zeroshot performance on Spanish and Greek. Results showed an F1 drop of roughly 15% compared with monolingual baselines, but the system still outperformed random and lexical baselines, confirming the viability of shared transformers.

ProjectionBased Transfer (2022) By exploiting the EuroParl parallel corpus, metaphor labels were projected from English to German and French. After a short finetuning stage on the projected data, the models achieved performance close to fully supervised German and French detectors, with an F1 within 5% of the monolingual upper bound.

Contrastive Multilingual Training (2023) A contrastive loss was added to encourage representations of metaphorical expressions to cluster together across languages while keeping literal expressions separate. The approach yielded consistent gains (37% absolute F1) for lowresource languages such as Lithuanian and Urdu.

Practical Guidelines for Building a CrossLingual Detector

1. Choose the Right Multilingual Backbone

For most applications, XLMR or multilingual T5 provide strong crosslingual alignment out of the box. If you need a lightweight solution, distilled versions (e.g., distilmBERT) reduce inference time with modest loss in accuracy.

2. Gather Parallel or Comparable Data

Even a modest amount of sentencealigned text (e.g., 100k sentences) can be sufficient for effective label projection. When parallel data is scarce, comparable corpora (e.g., news articles covering the same events) can be aligned using sentence similarity measures.

3. Perform Careful Alignment

Word alignment tools such as fast_align or eflomal work well for European languages. For more divergent language pairs, neural alignment models (e.g., Awesome-align) give higher precision.

4. Mitigate Noise in Projected Labels

Projection introduces errors because literal sentences can be misaligned with metaphorical ones. Strategies to clean the data include:

• Discarding lowconfidence alignments.
• Applying a surfacelevel heuristic (e.g., keep only words flagged as metaphorical by a bilingual dictionary).
• Using a small manually verified seed set to retrain a classifier that filters noisy examples.

5. Leverage FewShot FineTuning

If you can afford to annotate 200500 sentences in the target language, the improvement is often dramatic. A mixedobjective finetuning (crossentropy + contrastive loss) works well in this regime.

6. Evaluate with a Balanced Test Set

Metaphor detection datasets are typically imbalanced (metaphorical instances are rarer). Ensure the test set contains a realistic proportion of metaphorical vs. literal sentences, and report both macroaveraged F1 and perclass scores.

Future Directions

• ZeroShot Prompting: Large multilingual language models (e.g., GPT4o) can be prompted to label metaphors without any finetuning. Systematic studies are needed to compare this approach with traditional transfer.
• Multimodal Metaphor Detection: Images, videos, and gestures often accompany metaphoric language. Aligning visual embeddings with multilingual text may improve detection, especially for cultural metaphors.
• CrossCultural Metaphor Resources: Building shared metaphor ontologies that encode culturespecific source domains could guide transfer learning and reduce false positives.

Conclusion

Crosslingual model transfer provides a practical pathway to extend metaphor detection to languages that lack large annotated corpora. By combining multilingual pretrained transformers, careful projection of annotations, and modest targetlanguage finetuning, researchers can achieve robust performance across diverse linguistic settings. Ongoing advances in multilingual representation learning and promptbased inference promise to further narrow the gap between highresource and lowresource languages, bringing nuanced figurativelanguage understanding to a truly global audience.

For further reading, see the following resources:

• Multilingual BERT for Metaphor Detection
• ProjectionBased Metaphor Transfer
• Contrastive Learning for CrossLingual Metaphor Detection