💬 NLP

Wisdom in Unity: The Role of Multilingual Training in Figurative Language Identification in Proverbs

This paper demonstrates that multilingual supervision, particularly when incorporating diverse figurative forms like Culture Specific and Moral/Advisory proverbs, significantly enhances figurative language identification performance and suggests a shift toward multidimensional, concept-level frameworks beyond metaphor-centric approaches.

Original authors: Rama Alomair, Remas Alsubaie, Walaa Saifalislam, Rima Alsonbul, Mona Alnajjar, Razan Aldossari, Haya Alibrahim, Abeer Aldayel

Published 2026-08-11

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Rama Alomair, Remas Alsubaie, Walaa Saifalislam, Rima Alsonbul, Mona Alnajjar, Razan Aldossari, Haya Alibrahim, Abeer Aldayel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to understand human jokes, riddles, and old sayings. This isn't just about translating words from one language to another; it's about understanding the hidden meaning behind them. In the world of computer science, this is called "figurative language." Think of it like a secret code where "it's raining cats and dogs" doesn't mean animals are falling from the sky, but that the rain is heavy. Proverbs are the ultimate test for these robots because they are short, packed with cultural wisdom, and often make no sense if you take them literally. For a long time, scientists tried to teach computers to spot these sayings using only one language at a time, like studying English jokes in an English-only classroom. But the real world is messy and multilingual, so researchers wanted to know: if we show a robot sayings in many different languages at once, does it get smarter? Does seeing a proverb in French, Japanese, and Arabic help it understand the proverb in English better?

This paper, titled "Wisdom in Unity," dives into that exact question using a massive collection of 742 proverb concepts translated into seven different languages (Arabic, English, French, German, Russian, Japanese, and Spanish), totaling over 6,700 examples. The researchers didn't just ask if a proverb is "figurative" or "literal"; they broke it down into four specific flavors of meaning: Metaphorical (using images or symbols), Moral/Advisory (giving a lesson or warning), Cause–Effect (linking an action to a result), and Culture-Specific (needing deep local knowledge to get the joke). They tested five different computer models, ranging from standard encoders to advanced "instruction-tuned" large language models (LLMs), to see how well they performed when trained with different amounts of these translated examples.

Here is what they discovered, and it's a bit like finding the perfect recipe for a multilingual smoothie. First, they found that you don't need to feed the robot every single translated proverb to get great results. In fact, feeding the models about 50% of the translated training data was enough to reach near-perfect performance. Adding more data beyond that point (going from 50% to 100%) didn't help much, suggesting that the models hit a "sweet spot" where they had learned enough to generalize.

Second, and perhaps most surprisingly, the type of proverb matters a lot. The researchers found that the "Culture-Specific" proverbs—the ones that are the hardest to understand because they rely on local history or religion—actually saw the biggest improvement when the models were trained with multilingual data. It's as if seeing the proverb in different cultural contexts helped the robot finally crack the code of what makes it unique. Meanwhile, the more common types, like metaphors, didn't get as big of a boost from the extra languages.

Finally, the study showed that no single "flavor" of proverb is enough on its own. The models performed best when they were trained on a mix of all four types. It turns out that while some models might be great at spotting moral lessons and others at spotting cause-and-effect, combining all these different perspectives creates the strongest overall understanding. The authors suggest that instead of just looking for metaphors, future AI needs to be trained to recognize these different, complementary layers of meaning to truly understand the wisdom hidden in our global collection of proverbs.

Technical Summary: Wisdom in Unity

Problem Statement
The interpretation of figurative language, particularly proverbs, remains a significant bottleneck for Natural Language Processing (NLP). Unlike literal text, the semantic essence of a proverb is often "non-compositional," meaning its significance cannot be derived solely from the individual meanings of its constituent words. While prior work has treated figurativeness as a linguistic phenomenon focusing on specific rhetorical devices (e.g., metaphors, similes, idioms), there is a lack of understanding regarding how translated multilingual supervision contributes to figurative language identification. Furthermore, existing approaches often rely on language-homogeneous training data or binary classification, lacking a nuanced framework to analyze how different dimensions of figurative meaning transfer across languages.

Methodology
The study addresses these gaps through a comprehensive evaluation of multilingual supervision on proverb identification using a dataset of 742 unique proverb concepts represented by 6,787 translated instances across seven languages: Arabic, English, French, German, Russian, Japanese, and Spanish.

Multidimensional Annotation Framework: Moving beyond binary figurative-vs-literal classification, the authors introduce a fine-grained, theory-grounded annotation framework based on Relevance Theory. This framework characterizes proverbs through four complementary, non-exclusive figurative forms:
- Metaphorical: Meaning conveyed through imagery, analogy, or symbolic entities.
- Moral/Advisory: Guidance, warnings, or ethical judgments.
- Cause–Effect: Links between actions and their outcomes.
- Culture-Specific: Meanings requiring cultural, religious, historical, or local knowledge.
  Annotations were performed on a ground-truth subset (83 concepts) by human annotators, achieving substantial agreement (Cohen's $\kappa$ up to 0.622). The remaining dataset was labeled using weak supervision models (XLM-RoBERTa for binary, E5 embeddings with Logistic Regression for multi-label).
Experimental Setup: The authors evaluated five models representing different architectural families: Logistic Regression (baseline), mBERT, XLM-RoBERTa Large, Aya-8B, and Qwen2.5-3B (instruction-tuned LLMs).
- Progressive Supervision (RQ1): Models were trained with fixed source-language data while progressively increasing the proportion of translated multilingual training data (0%, 10%, 50%, and 100%).
- Instance-Selection Ablation (RQ2): To isolate the impact of specific figurative forms, the authors trained models on fixed-size subsets (600 instances) containing only specific figurative forms (e.g., only Metaphorical, only Culture-Specific) combined with the source data.

Key Results

Impact of Multilingual Supervision: Increasing multilingual supervision consistently improves performance across all models and languages. However, the gains exhibit diminishing returns; approximately 50% of translated multilingual training data is sufficient to achieve near-optimal performance, with smaller gains observed between 50% and 100% supervision.
Model Performance: Instruction-tuned LLMs (Aya-8B, Qwen2.5-3B) generally achieve larger improvements with small amounts of additional supervision compared to encoder-based models. Qwen2.5-3B achieved the highest Macro-F1 on four of the seven languages (Arabic, French, German, Russian), while mBERT performed best on Spanish and XLM-R Large on Japanese.
Differential Benefits of Figurative Forms: The study found that different figurative forms benefit unevenly from multilingual supervision.
- Culture-Specific: Despite being the least frequent form (approx. 6-14% of data) and having the lowest overlap with other forms, it exhibits the largest performance gains under multilingual supervision.
- Complementarity: While individual forms provide specific signals (e.g., Moral/Advisory benefits Qwen2.5-3B, Cause-Effect benefits XLM-R Large), combining all diverse figurative forms yields the strongest overall performance.
Preservation of Forms: The prevalence of the four figurative forms remains relatively stable across the seven translated languages, suggesting that the distribution of figurative meaning is generally preserved in translation, even if individual instance-level interpretations may vary.

Significance and Contributions
The paper claims three primary contributions:

Empirical Evaluation of Supervision Levels: It provides a systematic evaluation of how progressively increasing multilingual supervision (0% to 100%) affects figurative identification across diverse model architectures and seven target languages.
Theoretical Framework for Analysis: It introduces a multidimensional annotation framework that moves beyond metaphor-centric taxonomies. This framework allows for a finer-grained analysis of how complementary dimensions of figurative meaning (Metaphorical, Moral, Cause-Effect, Culture-Specific) transfer across languages.
Insights on Transferability: The findings demonstrate that multilingual supervision benefits figurative forms unevenly. Specifically, the study highlights that "culture-specific" knowledge, often considered a barrier in cross-lingual transfer, shows the most significant performance gains when multilingual data is introduced.

Conclusion
The authors argue that these findings motivate a shift in multilingual figurative identification research. Rather than treating figurative language as a single phenomenon or relying solely on metaphor-centric approaches, future frameworks should adopt concept-level, multidimensional models. These models should explicitly account for complementary forms of figurative meanings that are representative of the specific context, acknowledging that different forms of figurative knowledge possess distinct transfer characteristics.

Technical Summary: Wisdom in Unity

More like this