Multi-Agent Translation of Wordplay with Contrastive Learning for CLEF JOKER 2025
A recent research article introduces an innovative method for translating puns from English into French, integrating large language models with specialized wordplay generation techniques. The proposed approach consists of three phases: initially, a baseline is established using various advanced large language models, guided by a new contrastive learning dataset; next, a chain-of-thought pipeline is utilized, incorporating phonetic-semantic embeddings; finally, a multi-agent generator-discriminator framework is applied to assess and regenerate puns with feedback. The main goal is to preserve the linguistic creativity and humor of the original text rather than merely translating it literally. This study tackles the distinctive difficulties of wordplay translation, which has historically posed challenges for both human translators and machine systems. The paper can be found on arXiv under the identifier 2507.06506, categorized as replace-cross, and is part of the CLEF JOKER 2025 Task 2, focusing on humor and wordplay translation.
Key facts
- Research proposes a novel approach for translating puns from English to French.
- Methodology combines large language models with specialized techniques for wordplay generation.
- Three-stage approach: baseline with frontier LLMs, guided chain-of-thought with phonetic-semantic embeddings, and multi-agent generator-discriminator framework.
- Uses a new contrastive learning dataset for feedback.
- Objective is to capture linguistic creativity and humor, not literal translation.
- Paper available on arXiv with identifier 2507.06506.
- Part of CLEF JOKER 2025 Task 2.
- Announcement type is replace-cross.
Entities
Institutions
- CLEF JOKER
- arXiv