On the Poor Robustness of Transformer Models in Cross-Language Humor Recognition

Humor is a pervasive communicative device; nevertheless, its portability from one language to another remains challenging for computer machines and even humans. In this work, we investigate the problem of humor recognition from a cross -language and cross-domain perspective, focusing on English and Spanish languages. To this aim, we rely on two strategies: the first is based on multilingual transformer models for exploiting the cross-language knowledge distilled by them, and the second introduces machine translation to learn and make predictions in a single language. Experiments showed that models struggle in front of the humor complexity when it is translated, effectively tracking a degradation in the humor perception when mes-sages flow from one language to another. However, when multilingual models face a cross-language scenario, exclusive between the fine-tuning and evaluation data lan-guages, humor translation helps to align the knowledge learned in fine-tuning phase. According to this, a mean increase of 11% in F1 score was observed when classi-fying English-written texts with models fine-tuned with a Spanish dataset. These results are encouraging and constitute the first step towards a computationally cross -language analysis of humor.