Sentiment analysis plays a crucial role in interpreting human interactions across different modalities. This paper addresses the Cross-Cultural Multilingual Humour Detection Sub-Challenge in the MuSe 2023 Multi-modal Sentiment Analysis Challenge, which aims to identify humor instances in cross-cultural situations using multimodal data. In this paper, we explore the unique window length processing of audio data to more accurately and fully capture humor information. Furthermore, we employ a data augmentation technique for text to bridge the gap between the train set and the test set. This augmentation involves converting German and English samples in the database to each other, as well as translating both languages into Chinese. These techniques aim to enhance the model's generalization across different languages and alleviate the scarcity of humor samples. For multimodal fusion, we use the late fusion technique applied to different Gated Recurrent Unit (GRU) models with modality-specific weights obtained through gradient descent. Experimental results demonstrate the effectiveness of our approach, achieving an Area Under the Curve (AUC) score of 0.872 on the MuSe-Humor test set. This performance ranked second place in the sub-challenge.
Reusing separate, pre-filled Key-Value (KV) Caches for multiple contexts has become a common practice in handling multi-context scenarios with Large Language Models. However, this leads to a lack of cross-attention mechanisms between contexts. To address this, we propose CatLLM, the first method that concatenates multiple contexts across requests offline to compensate for this deficiency. Specifically, during offline processing, CatLLM identifies contexts that severely lack cross-attention by incorporating the weighted inner products of Q and K vectors from tokens in an un-concatenated context into an equivalently transformed weighted formulation for concatenated Q and K inner products. This yields a weighting corresponding to the output vector difference, which can then be used to identify contexts with severe cross-attention deficiencies and concatenate them into a single context for KV Cache computation. Experimental results show that, compared to the baseline of separate caching (i.e., no concatenation), fully concatenating all contexts improves the F1 score by 6%. Meanwhile, the proposed method reduces the number of contexts requiring caching from 10 to 7 while achieving a 3% F1 score, thereby maximizing performance improvement while minimizing the degree of context compression.