DPP: A Dual-Phase Processing Method for Cross-Cultural Humor Detection

Multimodal humor detection has become an active research area in the field of artificial intelligence. This paper presents the solution for the MuSe-Humor sub-challenge of cross-cultural humor detection. The goal of the MuSe-Humor sub-challenge is to predict whether humor exists in a given dataset, which includes data from various modalities such as text, audio, and video. The training data consists of German recordings and their transcriptions, while the test data is in English. This cross-cultural testing introduces new challenges, distinguishing it from ordinary multimodal humor detection tasks. To address this issue, we propose a method called Dual-Phase Processing (DPP). The proposed method first preprocesses the text data using a large language model to extract more effective features for humor detection from the raw data. It also partially addresses cross-cultural differences through bilingual annotation. Finally, by applying composite temporal smoothing to the results after decision-level fusion of the three modalities, the accuracy of humor prediction is greatly improved. The experimental results on the official test set demonstrate the effectiveness of our model. Our model achieves an impressive AUC score of 0.9366 on the test set, far surpassing the baseline and securing the first-place ranking.