Multi-level Multi-task representation learning with adaptive fusion for multimodal sentiment analysis

Multimodal sentiment analysis is an active task in multimodal intelligence, which aims to compute the user’s sentiment tendency from multimedia data. Generally, each modality is a specific and necessary perspective to express human sentiment, providing complementary and consensus information unavailable in a single modality. Nevertheless, the heterogeneous multimedia data often contain inconsistent and conflicting sentiment semantics that limits the model performance. In this work, we propose a Multi-level Multi-task Representation Learning with Adaptive Fusion (MuReLAF) network to bridge the semantic gap among different modalities. Specifically, we design a modality adaptive fusion block to adjust modality contributions dynamically. Besides, we build a multi-level multimodal representations framework to obtain modality-specific and modality-shared semantics by the multi-task learning strategy, where modality-specific semantics contain complementary information and modality-shared semantics include consensus information. Extensive experiments are conducted on four publicly available datasets: MOSI, MOSEI, SIMS, and SIMSV2(s), demonstrating that our model exhibits superior or comparable performance to state-of-the-art models. The achieved accuracies are 86.28%, 86.07%, 84.46%, and 82.78%, respectively, showcasing improvements of 0.82%, 0.84%, 1.75%, and 1.83%. Further analyses also indicate the effectiveness of our model in sentiment analysis.