基于对比学习与大语言模型增强的多模态方面级情感分析模型OA

[Purpose/Significance]This study addresses challenges such as data sparsity and class imbalance in the field of Multimodal Aspect-Based Sentiment Analysis(MABSA),and explores the application and performance of large lan-guage models in MABSA tasks.[Methods/Process]The paper proposed a multimodal aspect-based sentiment analysis model named HLCL-GLM4,which integrates data enhancement using large language models and HiLo-attention-based contrastive learning.The model leveraged ChatGLM4-Flash to perform data enhancement,extracted textual and visual fea-tures using BART embeddings and Faster R-CNN,respectively,and encoded image features with HiLo attention mecha-nism.The model employed a self-supervised contrastive learning strategy to facilitate multimodal feature learning and fusion,which improves sample diversity and the expressiveness of sentiment semantics.[Result/Conclusion]Experimental results show that HLCL-GLM4 achieves superior performance on both the Twitter-15 and Twitter-17 datasets.Specifi-cally,compared with the best baseline model,HLCL-GLM4 improves the F1-score by 1.6%on Twitter-15 and by 0.8%on Twitter-17.