基于多头自注意力机制与MLP-Interactor的多模态情感分析

林宜山,左景,卢树华

Multimodal sentiment analysis based on multi-head self-attention mechanism and MLP-Interactor

Yishan LIN,Jing ZUO,Shuhua LU

表 4 在CMU-MOSI数据集上的消融实验结果

Tab.4 Result of ablation experiment on CMU-MOSI dataset

方法 MAE corr A2 A7 F1 BERT 0.799 0.746 80.80/82.87 40.77 80.91/82.90 DeBERTa 1.154 0.486 66.96/68.06 28.86 66.93/67.93 RoBERTa 0.664 0.824 83.63/85.65 45.53 83.65/85.64 DistilBERT 0.754 0.768 81.40/83.54 40.48 81.48/83.55 ALBERT 0.928 0.670 76.93/79.00 34.67 77.10/79.08 w/o Intra-Modality Interaction 0.613 0.851 86.90/89.10 47.92 86.90/89.10 w/o MLP-Interactor 0.590 0.868 85.57/87.54 50.00 85.66/87.59 w/o audio 0.635 0.853 86.46/88.16 44.79 86.48/88.15 w/o video 0.583 0.861 87.35/89.42 49.55 87.45/89.47 w/o text 1.460 0.052 45.95/47.47 14.58 50.86/52.50 本文模型 0.575 0.868 87.60/89.60 52.23 87.70/89.60