This paper presents a pun-focused translation system developed for the JOKER-2025 Task 2 competition on English-to-French pun translation. The system combines mBART-50 fine-tuning, LoRA adapter training, and a novel phoneme-aware forced-decoding strategy guided by a specialized pun dictionary. The end-to-end pipeline encompasses robust pun detection and tagging, oversampling-based data augmentation, phoneme transcription, and sentiment feature enrichment, enabling comprehensive capture of the layered meanings and playful ambiguity inherent in puns. Multiple model configurations and decoding schemes were evaluated, with the best configuration achieving a BLEU score of 32.82 and ranking 22nd overall. These findings underscore the efectiveness of incorporating phonetic cues and lexical constraints to enhance both faithfulness and humor preservation in pun translation across languages.
The automatic translation of humor remains one of the most persistent challenges in natural language
processing (NLP), especially when it comes to wordplay such as puns. Recent studies [
Among humorous forms, puns stand out for their intricate layering of meanings through phonological
resemblance, lexical ambiguity, or syntactic manipulation. This dual (or multiple) reading efect often
exploits language-specific structures, making literal translation inefective or outright impossible [
In response to these challenges, shared tasks like JOKER [
Recent developments in multilingual sequence-to-sequence modeling, most notably the mBART-50 architecture combined with parameter-eficient adaptation techniques such as LoRA adapters [ 7, 8] have unlocked powerful domain-specific translation capabilities. Complementing these modeling advances, work on lexically constrained decoding [9] and full phoneme-level pipelines [10] has demonstrated that explicit control over the generation process can help retain critical structural elements in challenging tasks, such as pun translation.
A variety of specialized methods for handling wordplay have since emerged: • Dual-attention architectures that localize and interpret puns via gloss pairs [11]. • GPT-3–based generators exploiting contextual ambiguity for creative pun synthesis [12]. • Enriched corpora annotated with keywords and explanatory notes to boost classification and generation quality [13]. • Unified frameworks addressing both homophonic and homographic puns [14]. • Grapheme-to-phoneme augmentation techniques preserving acoustic wordplay signals [15]. • Large-model evaluations exposing “lazy” pun reproduction and motivating new metrics [16]. While these approaches advance semantic disambiguation, data balancing, and lexical enforcement, few integrate phonetic features directly into the decoding mechanism itself.
This paper proposes an end-to-end pipeline for English-to-French pun translation that unites mBART50 fine-tuning (with optional LoRA adapters), systematic pun detection and markup, explicit G2P-based phoneme transcription, and a phoneme-aware forced-decoding strategy guided by a curated pun dictionary. By enriching inputs with sentiment and phonetic embeddings, experiments show substantial gains in preserving both the humorous intent and the phonological wordplay of source puns. This paper is structured as follows: Section 2 describes the methodology, including dataset description, pun detection and tagging, data augmentation, feature enrichment, and model training. Section 3 presents performance results and qualitative analysis of the pipeline variants. Finally, Section 4 ofers conclusions and outlines directions for future work.
This section describes the architecture of the English-to-French pun translation system developed for Task 2 of the JOKER-2025 competition. As shown in Figure 1, the workflow comprises five main stages: • Data Preprocessing: Automated detection of puns in the source text and insertion of explicit markup to guide translation. • Feature Engineering: Enrichment of input data with sentiment labels and phonemic transcriptions to capture both emotional tone and sound-based wordplay. • Model Training: Comparison of two complete fine-tuning training strategies versus the LoRA adaptation that is eficient in parameterizing applied to the multilingual model mBART-50. • Inference: Application of a custom forced decoding algorithm, guided by a pun-specific lexicon, to ensure the preservation of key wordplay elements during translation. • Evaluation: Performance assessment using both automatic metrics and manual evaluation criteria defined by the JOKER track, with an emphasis on maintaining the intended humorous efect.
The dataset used in this work was released as part of the JOKER 2025 shared task on wordplay translation
[
Pun Detection & Tagging
Data Augmentation (3x oversampling)
Cleaning
Sentiment Analysis Phoneme Transcription
Training and Model
Selection mBART-50 Fine-Tuning LoRA Adapter
Trainig Tokenizer
Evaluation each entry formatted as a JSON object containing: id_en (unique identifier), en (source English text containing wordplay), and fr (French translation).
A distinctive characteristic of the training data is the presence of multiple French translations for each English pun, reflecting diferent approaches to wordplay preservation. The dataset contains training examples similar to those provided in previous editions, ofering diverse examples of pun translation strategies.
The test dataset follows the same format but contains only the original English text (en) and identifier (id_en), requiring systems to generate French translations that preserve both semantic meaning and humorous efect. The evaluation prioritizes wordplay preservation over traditional translation metrics, following Delabastita’s pun translation strategy. Table 1 shows the composition of the dataset provided by the competition organizers.
The preprocessing pipeline implements a systematic approach to prepare the pun translation dataset for model training. As illustrated in Figure 1, the preprocessing stage consists of two sequential operations designed to enhance wordplay structure representation and address dataset imbalances inherent in pun translation tasks.
English puns are automatically identified by matching the input text against a curated lexicon, and each matched occurrence is wrapped with special delimiters to guide the translator, as shown in Algorithm 1.
• Each input sentence is iterated over and scanned against the English to French pun dictionary. • On the first case-insensitive match, <PUN_EN> is injected before and after the English pun. • During training the French side can receive <PUN_FR> around the mapped French entry. • This explicit markup focuses the downstream model on the wordplay span without disturbing the surrounding context.
Algorithm 1 Pun Detection and Tagging Require: List of sentences {}, pun lexicon ℒ = {(, )} where is English pun, its French equivalent Ensure: Tagged sentences with ⟨PUN_EN⟩, ⟨PUN_FR⟩ 1: for each sentence in input do 2: tagged ← 3: for each (, ) in ℒ do 4: if lower() occurs in lower() then 5: Mark first match: tagged ← replaceOnce( tagged, , ⟨PUN_EN⟩ ⟨PUN_EN⟩) 6: break ◁ Only tag one pun per sentence 7: end if 8: end for 9: Output tagged 10: end for
Tagging Example Original: A skier retired because he was going downhill.
Tagged:
A skier retired because he was going <PUN_EN> downhill <PUN_EN>.
Targeted oversampling is applied to mitigate the imbalance between punny and non-punny examples. Each pun-containing instance is replicated three times in the training set. This ensures a balanced representation of wordplay cases during learning, following established minority-class techniques to improve the model’s ability to handle scarce patterns.
As illustrated in Figure 2, the augmentation process operates systematically through four sequential stages.
• Pun Identification: Identify all examples containing at least one detected pun. • Data Augmentation: Create three additional copies of each example containing puns. • Dataset Consolidation: Combine original and augmented examples into a unified training set. • Dataset Shufling : Shufle the dataset using a fixed random seed for reproducibility.
Train Dataset
Pun Identification
Data Augmentation
Dataset Consolidation
Dataset Shuffling
The augmentation process creates a more balanced training distribution, allowing the fine-tuned model to develop robust representations for both standard translation patterns and specialized wordplay mechanisms. Without this rebalancing, the model would predominantly optimize for literal translation patterns, potentially failing to preserve wordplay in test scenarios.
A comprehensive feature engineering approach is implemented to incorporate linguistic characteristics relevant to wordplay translation. This stage enriches the preprocessed dataset with semantic and phonetic features designed to capture emotional and acoustic patterns that may influence preservation strategies for puns.
Each training instance is augmented with sentiment classification to capture the emotional context of humorous content. Sentiment analysis is performed by a pre-trained transformer model (cardifnlp/twitterroberta-base-sentiment-latest) fine-tuned on social media data [ 17], providing robust performance for informal and creative text typical in pun datasets.
The sentiment classifier assigns one of three labels to each original English sentence: POSITIVE, NEGATIVE, or NEUTRAL. This categorical sentiment information is stored in a new field in the dataset, providing the translation model with explicit emotional context that can guide the preservation of the appropriate tone in the target language.
Automatic phonemic transcription is implemented using a grapheme-to-phoneme (G2P) conversion system [15] to capture phonetic patterns relevant to sound-based wordplay. The G2P model converts English text to ARPAbet phonetic notation, providing explicit acoustic representations that may inform the model about sound-based pun mechanisms. The phonemic transcription is stored in a new field, enabling the model to access pronunciation patterns that are often crucial for wordplay preservation, particularly for puns based on homophones or rhymes. Each training record contains a unique identifier, the English pun, and one or more human-verified French translations. For feature enrichment, each sentence is paired with its phonemic transcription in ARPAbet notation to highlight sound-based wordplay structures. Algorithm 2 summarizes the phoneme extraction process.
Algorithm 2 Detailed Feature Enrichment Process Require: Dataset of pun sentences with unique IDs and English text Ensure: The same dataset with added sentiment and phoneme annotations 1: for each record in the dataset do 2: Extract the English text 3: Preprocess the text: • Lowercase normalization • Removal of extra whitespace or unwanted characters 4: 5:
Predict the sentiment label (POSITIVE, NEGATIVE, NEUTRAL) using a pre-trained sentiment classification model
Generate the phoneme transcription: • Tokenize the text into words • For each word, obtain its ARPAbet phoneme sequence • Reconstruct the full sentence phoneme string, preserving punctuation 6: Add the sentiment label and phoneme string to the record 7: Leave the translation field empty for future prediction 8: end for 9: return Enriched dataset in JSON format with fields: id_en, en, sent, phon, fr Each input pun undergoes careful preprocessing, sentiment classification, and phoneme conversion to enrich its representation for training. This step enables the translation model to capture the emotional tone and sound-based ambiguities inherent in puns. Algorithm 2 describes this process in detail, and an example output is presented below.
Example of a Fully Enriched Record
id_en: en_1 en: Save the whales, spouted Tom. sent: POSITIVE phon: S EY1 V DH AH0 W EY1 L Z , fr: Sauvez les baleines, souffla Tom.
S P AW1 T AH0 D
Two distinct fine-tuning strategies for pun translation are evaluated, comparing complete parameter optimization against parameter-eficient adaptation techniques. Both approaches utilize mBART-50 (facebook/mbart-large-50-many-to-many-mmt) as the base multilingual transformer [18], configured for English-to-French translation with source language token en_XX and target language token fr_XX.
Algorithm 3 details the training procedure for both full fine-tuning and LoRA adaptation. Algorithm 3 Model Training for Pun Translation Require: Enriched training data, pre-trained mBART-50 model Ensure: the Fine-tuned model is ready for pun translation 1: Initialize the mBART-50 model and tokenizer 2: Define the objective: maximize translation accuracy while preserving wordplay 3: Tokenize sentences and truncate to maximum length 4: Create mini-batches and set optimizer and learning rate scheduler 5: if Full fine-tuning then 6: Update all model weights 7: else 8: Freeze base weights and train only low-rank adapters (LoRA) 9: end if 10: for each training epoch do 11: Compute loss on each batch 12: Backpropagate gradients and update weights 13: Periodically evaluate the validation set 14: end for 15: Select best checkpoint based on validation performance 16: return Trained model for inference
The first approach implements complete model fine-tuning, updating all parameters of the mBART50 architecture during training. This method provides maximum model expressivity by allowing unrestricted parameter updates across all transformer layers, attention mechanisms, and feed-forward networks.
The training process employs the Seq2SeqTrainer framework with the hyperparameter configuration shown in Table 2.
The complete fine-tuning approach enables comprehensive adaptation to the pun translation domain, allowing the model to learn specialized linguistic patterns for wordplay preservation across all network parameters. This method provides maximum learning capacity by updating the entire transformer architecture.
The second approach implements Low-Rank Adaptation (LoRA) [7], a parameter-eficient fine-tuning technique that freezes the pre-trained model weights and trains only low-rank decomposition matrices inserted into attention layers. This method significantly reduces the number of trainable parameters while maintaining competitive performance.
The LoRA configuration targets attention projection matrices with the settings detailed in Table 3.
The adapter training maintains the same core hyperparameters as full fine-tuning but operates on a drastically reduced parameter space. The training configuration turns of intermediate checkpointing and external logging for eficiency, thereby concentrating computational resources on the adapter optimization process.
Both training strategies incorporate the enhanced dataset with pun tagging and linguistic features, enabling comparison between maximum parameter flexibility and eficient domain adaptation for specialized translation tasks. The following boxed example illustrates how the trained model processes an enriched input step by step. It demonstrates how tokenization, phoneme cues, sentiment guidance, and forced decoding work together to generate a pun-preserving French translation.
Step-by-Step Example: From Enriched Input to Final Translation Input Record: id_en: en_1 en: Save the whales, spouted Tom. sent: POSITIVE phon: S EY1 V DH AH0
W EY1 L Z , Model Process:
S P AW1 T AH0 D 1. Tokenization: The English sentence is tokenized with special language tokens (e.g., en_XX, fr_XX). 2. Encoding: The tokenized input is converted into embeddings using mBART-50’s encoder layers. 3. Forced Decoding: During generation, the decoder applies lexical constraints guided by the pun dictionary and phoneme cues. 4. Sentiment Guidance: The model conditions the output tone to match the predicted sentiment (here: POSITIVE).
5. Output: The decoder generates a French pun translation that preserves the wordplay. Final Translation: fr: Sauvez les baleines, souffla Tom.
The JOKER-2025 Task 2 competition for English-to-French pun translation was addressed using a progressive pipeline comprising three configurations: a fully fine-tuned system and two parametereficient LoRA variants. Table 4 summarizes the oficial BLEU scores assigned during the automatic evaluation phase.
The results indicate that the Full Fine-tuning model substantially outperformed both LoRA-based variants, achieving approximately 4-5 points higher BLEU. It suggests that, for pun translation, extensive parameter updates allow the model to learn better the subtle interplay of lexical form and multiple senses typical of wordplay.
However, as established in the JOKER evaluation framework, BLEU alone cannot reliably measure the success of pun translation, which requires preserving both semantic coherence and creative linguistic distortions. Therefore, all submitted systems are additionally evaluated by human experts using criteria including lexical field preservation, pun form retention, appropriateness of style shift, humor maintenance, and the detection of syntactic or lexical inaccuracies. Ultimately, the final ranking prioritizes the number of translations that successfully transfer the original wordplay mechanisms.
Example of Successful Pun Preservation (Full Fine-tuning) English: “A skier retired because he was going downhill.” Phonemes: AH0 S K IY1 ER0 R IH0 T AY1 ER0 D B IH0 K AO1 Z IY1 W AA1 Z G OW1 IH0 NG D AW1 N HH IH1 L French (Full Fine-tuning): “Le skieur a pris sa retraite car il n’arrivait plus à remonter la pente.” Analysis: The model correctly preserves the double sense of “going downhill” (literal skiing + ifgurative decline) by using the idiomatic French expression “remonter la pente”, demonstrating strong handling of layered meaning and phonetic nuance.
Example of Failed Pun Preservation (Baseline LoRA) English: “Someone once accused me of stealing an old, rare, valuable stamp, and I philately denied it.” Phonemes: S AH1 M W AH1 N AH1 N S AH0 K Y UW1 Z D M IY1 AH0 V S T IY1 L IH0 NG AE1 N OW1 L D R EH1 R V AE1 L Y UW0 B AH0 L S T AE1 M P AE1 N D AY1 F IH0 L AE1 T L IY0 D IH0 N AY1 D IH0 T French (Baseline LoRA): “Quelqu’un m’a accusé d’avoir volé un timbre ancien et rare, et je l’ai nié sans hésiter.” Analysis: Here, the pun on “philately” (which plays on “flatly denied”) is entirely lost. The baseline LoRA output conveys a literal denial, missing both the wordplay and the sound-based twist. It shows LoRA’s limits without explicit phoneme constraints.
Combining automatic and qualitative results, the team achieved 5th place overall, with the highest BLEU of 32.82 for the Full Fine-tuning system. It confirms that extensive parameter updates yield more robust pun preservation. In contrast, the Baseline and Refined LoRA models, despite being computationally lighter, often fall back on literal phrasing, primarily when the pun hinges on phonetic similarity or double meanings. The contrast demonstrates that creative phenomena, such as puns, demand models to learn deeper associations across syntax, semantics, and phonology, which lightweight adapters alone may fail to capture fully.
Therefore, while BLEU provides an initial signal, manual expert evaluation remains crucial for fair ranking, as it verifies whether both the form and sense of the original wordplay survive translation. These insights support the adoption of hybrid approaches, which combine eficient adaptation with advanced decoding and post-editing to bridge the gap between automatic translation and human-like linguistic creativity.
Our system Full Fine-tuning achieved the highest BLEU (32.82, 22nd place) by fully updating the mBART50 weights, which allowed it to capture both the semantic and phonetic layers of puns (e.g. ’going downhill’remonter la pente). However, this approach incurs significant training and inference costs and still occasionally falters in out-of-vocabulary or heavily constrained puns. In contrast, the two LoRA variants (BLEU 28.17 and 27.57) required far fewer resources but consistently defaulted to literal renderings when phonetic or lexical cues were subtle (e.g., ’philately denied’plain denial), exposing their limited capacity to model creative wordplay.
These results underscore two key lessons: First, BLEU, while useful, only partially reflects the success of humor translation, as high overlap does not guarantee the preservation of wordplay form or tone shift. Second, manual expert evaluation remains indispensable to assess lexical field preservation, stylistic shifts, and the maintenance of humorous efect. Only human judgments can reliably rank systems by their ability to transfer both the form and the sense of puns across languages.
Future directions include: • Dynamic Phoneme Constraints: integrate on-the-fly phoneme matching to handle unseen puns. • Pun-Aware Metrics: develop semi-automatic measures that correlate with human judgments of wordplay. • Adapter Enhancements: augment lightweight LoRA with explicit phoneme–grapheme alignment layers to narrow the quality gap. • Cross-Domain Extension: apply and evaluate the pipeline on other language pairs and humor forms (idioms, jokes).
The authors would like to acknowledge the support provided by the master’s degree scholarship program in engineering at the Universidad Tecnologica de Bolivar (UTB) in Cartagena, Colombia.
During the preparation of this work, the author(s) used Grammarly in order to: Grammar and spelling check. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [7] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adaptation of large language models, 2021. URL: https://arxiv.org/abs/2106.09685. arXiv:2106.09685. [8] T. Dettmers, A. Pagnoni, A. Holtzman, L. Zettlemoyer, Qlora: Eficient finetuning of quantized llms, 2023. URL: https://arxiv.org/abs/2305.14314. arXiv:2305.14314. [9] S. Braun, O. Vasilyev, N. Iskender, J. Bohannon, Does summary evaluation survive translation to other languages?, in: M. Carpuat, M.-C. de Marnefe, I. V. Meza Ruiz (Eds.), Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Association for Computational Linguistics, Seattle, United States, 2022, pp. 2425–2435. URL: https://aclanthology.org/2022.naacl-main.173/. doi:10. 18653/v1/2022.naacl-main.173. [10] X. Zhao, E. Durmus, D.-Y. Yeung, Towards reference-free text simplification evaluation with a BERT Siamese network architecture, in: A. Rogers, J. Boyd-Graber, N. Okazaki (Eds.), Findings of the Association for Computational Linguistics: ACL 2023, Association for Computational Linguistics, Toronto, Canada, 2023, pp. 13250–13264. URL: https://aclanthology.org/2023.findings-acl.838/. doi:10.18653/v1/2023.findings-acl.838. [11] S. Liu, M. Ma, H. Yuan, J. Zhu, Y. Wu, M. Al-Ajarmah, A dual-attention neural network for pun location and using pun-gloss pairs for interpretation, 2021. doi:10.48550/arXiv.2110.07209. [12] A. Mittal, Y. Tian, N. Peng, AmbiPun: Generating humorous puns with ambiguous context (2022) 1053–1062. URL: https://aclanthology.org/2022.naacl-main.77/. doi:10.18653/v1/2022. naacl-main.77. [13] J. Sun, A. Narayan-Chen, S. Oraby, A. Cervone, T. Chung, J. Huang, Y. Liu, N. Peng, ExPUNations: Augmenting puns with keywords and explanations (2022) 4590–4605. URL: https://aclanthology. org/2022.emnlp-main.304/. doi:10.18653/v1/2022.emnlp-main.304. [14] Y. Tian, D. Sheth, N. Peng, A unified framework for pun generation with humor principles (2022) 3253–3261. URL: https://aclanthology.org/2022.findings-emnlp.237/. doi: 10.18653/v1/ 2022.findings-emnlp.237. [15] A. Pine, P. William Littell, E. Joanis, D. Huggins-Daines, C. Cox, F. Davis, E. Antonio Santos, S. Srikanth, D. Torkornoo, S. Yu, G2P rule-based, index-preserving grapheme-to-phoneme transformations, in: S. Moeller, A. Anastasopoulos, A. Arppe, A. Chaudhary, A. Harrigan, J. Holden, J. Lachler, A. Palmer, S. Rijhwani, L. Schwartz (Eds.), Proceedings of the Fifth Workshop on the Use of Computational Methods in the Study of Endangered Languages, Association for Computational Linguistics, Dublin, Ireland, 2022, pp. 52–60. URL: https://aclanthology.org/2022.computel-1.7/. doi:10.18653/v1/2022.computel-1.7. [16] Z. Xu, S. Yuan, L. Chen, D. Yang, “a good pun is its own reword”: Can large language models understand puns?, in: Y. Al-Onaizan, M. Bansal, Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Miami, Florida, USA, 2024, pp. 11766–11782. URL: https://aclanthology.org/2024. emnlp-main.657/. doi:10.18653/v1/2024.emnlp-main.657. [17] D. Loureiro, F. Barbieri, L. Neves, L. Espinosa Anke, J. Camacho-collados, TimeLMs: Diachronic language models from Twitter, in: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Association for Computational Linguistics, Dublin, Ireland, 2022, pp. 251–260. URL: https://aclanthology.org/2022.acl-demo.25. doi:10.18653/ v1/2022.acl-demo.25. [18] Y. Tang, C. Tran, X. Li, P.-J. Chen, N. Goyal, V. Chaudhary, J. Gu, A. Fan, Multilingual translation with extensible multilingual pretraining and finetuning, arXiv preprint arXiv:2008.00401 (2020).