Researchers tackle pun translation using AI agents and phonetic embeddings

Translating jokes and puns between languages has historically stumped both professional translators and machine translation systems alike. A research team has developed a novel framework that uses advanced AI techniques to preserve the humor and creativity of wordplay when converting text from English to French, moving beyond simple word-for-word translation.

Three-Stage Methodology#

The approach unfolds in three sequential phases. Initially, researchers established a foundation by fine-tuning multiple state-of-the-art language models using feedback derived from a newly created contrastive learning dataset. Next, they deployed a guided chain-of-thought process that integrates both phonetic and semantic embeddings to guide pun generation. Finally, they built a multi-agent system where one component generates candidate translations while another evaluates and rejects weak attempts, iteratively improving results through this feedback loop.

Balancing Semantics and Sound#

The core innovation centers on capturing linguistic creativity rather than replicating source vocabulary. Rather than attempting literal translation, the system learns to recognize the interplay between semantic ambiguity and phonetic similarity—the twin pillars of effective wordplay. By combining phonetic analysis with meaning-based embeddings, the framework helps models understand what makes a pun work, then regenerate that same effect in the target language.

Competition Results#

The research was submitted to CLEF JOKER 2025 Task 2, an international competition focused on humor translation. The team's submissions achieved first and second place, with evaluations performed by native French speakers who assessed the quality and authenticity of the translated wordplay. This validates the approach's effectiveness at the intersection of translation studies and computational linguistics.