MSFNet: modality smoothing fusion network for multimodal aspect-based sentiment analysis

Multimodal aspect-based sentiment classification (MABSC) aims to determine the polarity of a given aspect in sentence by combining text and image information. Although corresponding sample are associated with information, their features represented distinct semantic spaces, creating substantial gap. Previous research focused primarily on identifying fusing aspect-level expressions different modalities while ignoring To this end, we propose novel analysis model named modality smoothing fusion network (MSFNet). In model, process unimodal aspect-aware via feature strategy partially bridge Then fuse smoothed deeply using multi-channel attention mechanism, obtain representation comprehensive representing capability, thereby improving performance classification. Experiments two benchmark datasets, Twitter2015 Twitter2017, demonstrate that our outperforms second-best 1.96% 0.19% terms Macro-F1, respectively. Additionally, ablation studies provide evidence supporting efficacy each proposed modules. We release code at: https://github.com/YunjiaCai/MSFNet .