多源知识引导的视觉置信度感知的多模态情感分析模型

彭菊红; 张智; 刘朋; 葛文慧; 柳陈; 廖凌鑫; 张凯

doi:10.11999/JEIT260063

多源知识引导的视觉置信度感知的多模态情感分析模型

doi: 10.11999/JEIT260063 cstr: 32379.14.JEIT260063

彭菊红^{1, 2},
张智^{1, 2},
刘朋^{1, 2},
葛文慧^{1, 2},
柳陈^{1, 2},
廖凌鑫^{1, 2},
张凯^3, ,

1.
湖北大学人工智能学院武汉 430062
2.
智能感知系统与安全教育部重点实验室武汉 430062
3.
武昌船舶重工集团有限公司武汉 430415

基金项目: 国家自然科学基金(62377009)

详细信息

作者简介:
彭菊红：女，副教授，研究方向为信息处理及人工智能算法研究

张智：男，硕士生，研究方向为多模态情感识别

刘朋：男，硕士生，研究方向为多模态融合算法及目标检测

葛文慧：女，硕士生，研究方向为多模态识别与跟踪

柳陈：女，硕士生，研究方向为多模态识别与跟踪

廖凌鑫：男，硕士生，研究方向为多模态情感识别

张凯：男，高级工程师，研究方向为自动化信息集成及优化算法研究

通讯作者:
张凯　29859491@qq.com

中图分类号: TN911.7; TP391
计量
- 文章访问数: 20
- HTML全文浏览量: 4
- PDF下载量: 2
- 被引次数: 0
出版历程
- 收稿日期: 2026-01-20
- 修回日期: 2026-04-20
- 录用日期: 2026-04-23
- 网络出版日期: 2026-05-13

A Multimodal Sentiment Analysis Model with Multi-source Knowledge guided Visual Confidence Perception

PENG Juhong^{1, 2},
ZHANG Zhi^{1, 2},
LIU Peng^{1, 2},
GE Wenhui^{1, 2},
LIU Chen^{1, 2},
LIAO Lingxin^{1, 2},
ZHANG Kai^{3
, ,}

1.
School of Artificial Intelligence, HuBei University, WuHan 430062, China
2.
Key Laboratory of Intelligent Perception Systems and Security, Ministry of Education, WuHan 430062, China
3.
Wuchang Shipbuilding Industry Group Co., Ltd, WuHan 430415, China

Funds: The National Natural Science Foundation of China (62377009)

摘要

摘要: 针对多模态情感分析中图文不一致、视觉模态置信度低、模态贡献不均衡等问题，该文提出一种多源知识引导的视觉置信度感知模型(MKVP)。首先，通过多源知识引导构建视觉置信度感知(VCP)模块，利用文本句法与细粒度属性先验对视觉特征进行质量评估，有效过滤图像中受环境干扰的冗余信息，并引导其特征分布。其次，为避免模型对文本模态产生过度依赖并平衡模态贡献，设计双流并行交互模块，通过跨模态注意力机制促进图文特征的深层对等交互，强化图像特征对文本语义的补充与修正作用。最后，引入全局门控融合机制，根据各模态的全局贡献程度动态调节融合权重，实现从单模态主导向多模态均衡协同决策的转变。在MVSA-Single, MVSA-Multiple及HFM数据集上识别准确率和F1分数分别达到了77.56%和76.70%、72.72%和70.66%、87.26%和86.78%，对比基线模型识别准确率和F1分数分别提升2.45%和3.68%、2.19%和2.21%、1.83%和1.91%。说明该模型能有效挖掘样本中图文之间更深层次的情感表达。
- 多模态情感识别 /
- 多模态融合 /
- 多源知识引导 /
- 视觉置信度
Abstract: Objective Multimodal sentiment analysis is often affected by visual noise from complex environments, image-text sentiment inconsistency, and imbalanced modality contributions. When all modalities are treated without distinction, visual noise can degrade model performance. A robust mechanism is therefore needed to evaluate visual confidence and filter redundant visual information. Methods A Multimodal Sentiment Analysis Model with Multi-Source Knowledge-guided Visual confidence Perception (MKVP) is proposed (Fig. 1). A multi-source knowledge guidance matrix is constructed using syntactic-dependency, sentiment-intensity, and aspect-focused operators (Fig. 2). Guided by this matrix, the Visual Confidence Perception (VCP) module measures semantic affinity and dynamically suppresses irrelevant visual noise (Fig. 3). A dual-stream parallel interaction module is then used to support deep cross-modal alignment, and a global gated fusion mechanism further adjusts the fusion weights of different modalities. Results and Discussions Extensive experiments are conducted on the MVSA-Single, MVSA-Multiple, and HFM datasets. The proposed MKVP model achieves accuracy and F1 scores of 77.56% and 76.70%, 72.72% and 70.66%, and 87.26% and 86.78%, respectively. Compared with the baseline models, the accuracy and F1 score are improved by 2.45% and 3.68%, 2.19% and 2.21%, and 1.83% and 1.91%, respectively (Table 3). Ablation studies show that each component contributes to performance, especially the VCP module, which filters visual noise and improves feature quality (Table 5). Feature-space visualization further confirms that the VCP module refines semantic representations by promoting clearer clustering of samples with the same sentiment polarity (Fig. 4). Case studies on mismatched image-text samples also verify the ability of the model to resolve cross-modal semantic conflicts (Table 6). Model-complexity analysis shows that MKVP maintains high computational efficiency and low inference latency (Table 8). Conclusions The proposed MKVP framework reduces the effects of visual noise and image-text sentiment inconsistency in multimodal sentiment analysis. By using multi-source knowledge to guide visual confidence perception and combining dual-stream interaction with dynamic gated fusion, the model learns robust sentiment representations from noisy multimodal data. This method provides an efficient and reliable solution for complex social media scenarios.
- Multimodal sentiment analysis /
- Multimodal fusion /
- Multi-source knowledge guidance /
- Visual confidence perception

HTML全文

图 1 MKVP模型框架

下载: 全尺寸图片幻灯片

图 2 多源知识引导矩阵构建图

下载: 全尺寸图片幻灯片

图 3 VCP模块结构图

下载: 全尺寸图片幻灯片

图 4 HFM经过VCP模块前后的t-SNE可视化图

下载: 全尺寸图片幻灯片

图 5 MVSA-Single数据集联合损失超参数敏感性分析

下载: 全尺寸图片幻灯片

表 1 数据集的统计信息

	Train	Val	Test	总计
MVSA-Single	3611	450	450	4511
MVSA-Multiple	13624	1700	1700	17024
HFM	19816	2410	2409	24635

下载: 导出CSV

表 2 实验参数设置

参数	MVSA-Single	MVSA-Multiple	HFM
批量大小	32	16	32
学习率	5E-5	2E-5	5E-5
迭代轮次	40
优化器	AdamW
嵌入维度	768
$ {\gamma }_{1} $, $ {\gamma }_{2} $, $ {\gamma }_{3} $	1.0, 1.0, 1.0	1.0, 2.0, 1.0	1.0, 1.0, 1.0
Dropout	0.3	0.5	0.3

下载: 导出CSV

表 3 MKVP与所有基线模型在3个数据集上的对比结果

形式	模型	MVSA-Single		MVSA-Multiple		模型	HFM
形式	模型	Acc	F1	Acc	F1	模型	Acc	F1
文本	CNN	0.6819	0.5590	0.6564	0.5766	CNN	0.8003	0.7572
	Bi-LSTM	0.7012	0.6506	0.6790	0.6790	Bi-LSTM	0.8190	0.7753
	BERT	0.7111	0.6970	0.6759	0.6624	BERT	0.8339	0.8326
	BiACNN	0.7036	0.6916	0.6847	0.6319	—	—	—
	TGNN	0.7034	0.6594	0.6967	0.6180	—	—	—
图像	ResNet-50	0.6467	0.6155	0.6188	0.6098	ResNet-50	0.7277	0.7138
图像	ViT	0.6378	0.6226	0.6194	0.6119	Vit	0.7309	0.7152
文本+图像	MultiSentiNet	0.6984	0.6984	0.6886	0.6811	Concat(3)	0.8174	0.7874
	MGNNS	0.7377	0.7270	0.7249	0.6934	D&R Net	0.8402	0.8060
	CLMLF	0.7511	0.7302	0.7053	0.6845	CLMLF	0.8543	0.8487
	GIGNN	0.7511	0.7333	0.7341	0.7096	GIGNN	0.8556	0.8487
	DIB	0.7605	0.7520	—	—	—	—	—
	MVCN	0.7606	0.7455	0.7207	0.7001	MVCN	0.8568	0.8523
	MFGFN	0.7622	0.7538	0.7082	0.6994	—	—	—
	D²R	0.7667	0.7559	0.7159	0.7085	D²R	0.8672	0.8625
	DTN	0.7711	0.7646	0.7070	0.6810	DTN	0.8697	0.8646
	MIGSIE	0.7640	0.7520	0.7272	0.7272	—	—	—
	MKVP	0.7756	0.7670	0.7272	0.7066	MKVP	0.8726	0.8678

下载: 导出CSV

表 4 多源知识注入位置对比结果

	MVSA-Single		MVSA-Multiple		HFM
	Acc	F1	Acc	F1	Acc	F1
MKVP-II	0.7356	0.7128	0.7165	0.6866	0.8701	0.8651
MKVP-IS	0.7289	0.7029	0.7200	0.6942	0.8651	0.8600
MKVP-GL	0.7511	0.7339	0.7159	0.6795	0.8622	0.8557
MKVP-LF	0.7489	0.7335	0.7165	0.6752	0.8552	0.8501
MKVP	0.7756	0.7669	0.7272	0.7066	0.8726	0.8678

下载: 导出CSV

表 5 消融实验结果

	MVSA-Single		MVSA-Multiple		HFM
	Acc	F1	Acc	F1	Acc	F1
w/o VCP	0.7522	0.7355	0.7078	0.6879	0.8564	0.8510
w/o JOL	0.7589	0.7428	0.7135	0.6890	0.8669	0.8523
w/o CMI	0.7611	0.7494	0.7106	0.6945	0.8597	0.8561
w/o GMF	0.7667	0.7558	0.7182	0.6987	0.8622	0.8552
w/o V-J	0.7511	0.7401	0.7006	0.6833	0.8497	0.8446
w/o V- J -C	0.7467	0.7361	0.7088	0.6814	0.8460	0.8413
w/o V-J-C-G	0.7422	0.7354	0.6917	0.6710	0.8447	0.8394
MKVP	0.7756	0.7669	0.7272	0.7066	0.8726	0.8678

下载: 导出CSV

表 6 案例对比结果

图像	图像标签	文本	文本标签	ResNet	BERT	MKVP-VCP	CLMLF	MKVP
	Pos	Harshad’s second Missionn ? @har1603 what did you do??? #appalled	Neu	Pos	Neg	Neu	Neu	Pos
	Neu	RT @crashspain: Wonderful Turner Field Tour today. So excited for baseball season. Thanks @Braves @BravesReddit	Pos	Pos	Pos	Pos	Pos	Pos
	Neg	#abandoned #ruins #haikyo #urbex	Neu	Neu	Neu	Neg	Neg	Neg
	Neu	RT@AUFAMILY: Good wins over evil as there are once again two lives oaks at Toomer’s Corner. War Eagle! #ToomersForever	Neg	Pos	Pos	Pos	Pos	Neg

下载: 导出CSV

表 7 文本抗噪声实验结果

噪声类型	噪声强度(%)	MVSA-Multiple		HFM
噪声类型	噪声强度(%)	Acc	F1	Acc	F1
Shuffle	10	0.7124	0.6820	0.8618	0.8622
	30	0.7065	0.6769	0.8502	0.8511
	50	0.6924	0.6620	0.8419	0.8428
无噪声	-	0.7272	0.7066	0.8726	0.8678

下载: 导出CSV

表 8 模型复杂度与效率计算结果

	Params(M)	FLOPs(G)	Time(ms)	MVSA-Single
MGNNS	73.78	48.41	14.24	0.7377
CLMLF	205.52	24.07	9.46	0.7511
D2R	345.54	25.39	36.21	0.7667
MKVP	175.11	22.48	13.66	0.7756

下载: 导出CSV

参考文献(23)

[1]	YUAN Yuan, LI Zhaojian, and ZHAO Bin. A survey of multimodal learning: Methods, applications, and future[J]. ACM Computing Surveys, 2025, 57(7): 167. doi: 10.1145/3713070.
[2]	LU Ming, DONG Zhiqiang, GUO Ziming, et al. A multi-modal sarcasm detection model based on cue learning[J]. Scientific Reports, 2025, 15(1): 10261. doi: 10.1038/s41598-025-94266-w.
[3]	ZHAO Kai, ZHENG Mingsheng, LI Qingguan, et al. Multimodal sentiment analysis—a comprehensive survey from a fusion methods perspective[J]. IEEE Access, 2025, 13: 64556–64583. doi: 10.1109/ACCESS.2025.3554665.
[4]	LIU Xinjing, LI Ruifan, YE Shuqin, et al. Multimodal aspect-based sentiment analysis under conditional relation[C]. The 31st International Conference on Computational Linguistics, Abu Dhabi, UAE, 2025: 313–323.
[5]	YU Bengong, LI Chenyue, and SHI Zhongyu. Multi-grained feature gating fusion network for multimodal sentiment analysis[J]. Knowledge and Information Systems, 2025, 67(8): 6879–6905. doi: 10.1007/s10115-025-02446-x.
[6]	HUANG Huiting, GONG Tieliang, HE Kai, et al. Robust multimodal sentiment analysis via double information bottleneck[J]. Information Fusion, 2026, 129: 103964. doi: 10.1016/j.inffus.2025.103964.
[7]	胡泽, 陈志南, 杨宏宇. 多源特征融合增强的虚假新闻检测方法[J]. 电子与信息学报, 2025, 47(8): 2919–2934. doi: 10.11999/JEIT250041. HU Ze, CHEN Zhinan, and YANG Hongyu. A fake news detection approach enhanced by multi-source feature fusion[J]. Journal of Electronics & Information Technology, 2025, 47(8): 2919–2934. doi: 10.11999/JEIT250041.
[8]	ZI Lingling, PAN Xiangkai, and CONG Xin. MFSC: A multimodal aspect-level sentiment classification framework with multi-image gate and fusion networks[J]. Electronics, 2024, 13(12): 2349. doi: 10.3390/electronics13122349.
[9]	YANG Xiaocui, FENG Shi, ZHANG Yifei, et al. Multimodal sentiment detection based on multi-channel graph neural networks[C]. The 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021: 328–339. doi: 10.18653/v1/2021.acl-long.28.
[10]	WANG Hongbin, REN Chun, and YU Zhengtao. Multimodal sentiment analysis based on cross-instance graph neural networks[J]. Applied Intelligence, 2024, 54(4): 3403–3416. doi: 10.1007/s10489-024-05309-0.
[11]	ZHONG Qihuang, DING Liang, LIU Juhua, et al. Knowledge graph augmented network towards multiview representation learning for aspect-based sentiment analysis[J]. IEEE Transactions on Knowledge and Data Engineering, 2023, 35(10): 10098–10111. doi: 10.1109/TKDE.2023.3250499.
[12]	KIM Y. Convolutional neural networks for sentence classification[C]. The 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), Doha, Qatar, 2014: 1746–1751. doi: 10.3115/v1/D14-1181.
[13]	ZHOU Peng, SHI Wei, TIAN Jun, et al. Attention-based bidirectional long short-term memory networks for relation classification[C]. The 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Berlin, Germany, 2016: 207–212. doi: 10.18653/v1/P16-2034.
[14]	LAI Siwei, XU Liheng, LIU Kang, et al. Recurrent convolutional neural networks for text classification[C]. The 29th AAAI Conference on Artificial Intelligence, Austin, USA, 2015: 2267–2273. doi: 10.1609/aaai.v29i1.9513.
[15]	HUANG Lianzhe, MA Dehong, LI Sujian, et al. Text level graph neural network for text classification[C]. The 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, 2019: 3444–3450. doi: 10.18653/v1/D19-1345.
[16]	XU Nan and MAO Wenji. MultiSentiNet: A deep semantic network for multimodal sentiment analysis[C]. The 2017 ACM International Conference on Information and Knowledge Management, Singapore, Singapore, 2017: 2399–2402. doi: 10.1145/3132847.3133142.
[17]	SCHIFANELLA R, DE JUAN P, TETREAULT J, et al. Detecting sarcasm in multimodal social platforms[C]. The 24th ACM International Conference on Multimedia, Amsterdam, Netherlands, 2016: 1136–1145. doi: 10.1145/2964284.2964321.
[18]	XU Nan, ZENG Zhixiong, and MAO Wenji. Reasoning with multimodal sarcastic tweets via modeling cross-modality contrast and semantic association[C]. The 58th Annual Meeting of the Association for Computational Linguistics, 2020: 3777–3786. doi: 10.18653/v1/2020.acl-main.349.
[19]	LI Zhen, XU Bing, ZHU Conghui, et al. CLMLF: A contrastive learning and multi-layer fusion method for multimodal sentiment detection[C]. Findings of the Association for Computational Linguistics: NAACL 2022, Seattle, USA, 2022: 2282–2294. doi: 10.18653/v1/2022.findings-naacl.175.
[20]	WEI Yiwei, YUAN Shaozu, YANG Ruosong, et al. Tackling modality heterogeneity with multi-view calibration network for multimodal sentiment detection[C]. The 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Toronto, Canada, 2023: 5240–5252. doi: 10.18653/v1/2023.acl-long.287.
[21]	CHEN Yifan, LI Kuntao, MAI Weixing, et al. D²R: Dual-branch dynamic routing network for multimodal sentiment detection[C]. The 2024 Conference on Empirical Methods in Natural Language Processing, Miami, USA, 2024: 3536–3547. doi: 10.18653/v1/2024.emnlp-main.207.
[22]	余本功, 石中玉. 深层注意力和两阶段融合的图文情感对比学习方法[J]. 计算机工程与应用, 2025, 61(3): 223–233. doi: 10.3778/j.issn.1002-8331.2309-0470. YU Bengong and SHI Zhongyu. Deep attention and two-stage fusion of image-text sentiment contrastive learning method[J]. Computer Engineering and Applications, 2025, 61(3): 223–233. doi: 10.3778/j.issn.1002-8331.2309-0470.
[23]	卜韵阳, 卜凡亮, 张志江. 多通道交互下全局语义信息增强的多模态情感分析[J]. 计算机工程与应用, 2025, 61(19): 137–146. doi: 10.3778/j.issn.1002-8331.2406-0376. BU Yunyang, BU Fanliang, and ZHANG Zhijiang. Multimodal sentiment analysis of global semantic information enhancement under multi-channel interaction[J]. Computer Engineering and Applications, 2025, 61(19): 137–146. doi: 10.3778/j.issn.1002-8331.2406-0376.