报告详情
MGLAN: Multimodal Grounded Language Alignment Network for Cross- Model Visual Reasoning and Scene Understanding
编号：113
访问权限：仅限参会人
更新：2026-11-10 16:48:06 浏览：8次
Online
报告开始：2026年11月21日 11:30（Asia/Kolkata）
报告时间：15min
所在会场：[S1] AI-Native Communication Networks [S1-1] AI-Native Communication Networks
摘要
Abstract Multimodal and grounded language processing is a frontier challenge of Natural Language Processing (NLP), where we need to interpret, align and reason on heterogeneous data modalities including vision, language and audio simultaneously. Current methods like ViLBERT, UNITER and OSCAR exhibit competitive performance on single tasks, but generally lack sufficient cross-modal semantic grounding, domain generalization ability and structured scene-level context integration. We present Multimodal Grounded Language Alignment Network (MGLAN), which combines vision feature extraction using a Vision Transformer (ViT) backbone, contextual language encoding with BERT-based representations, and also includes a rich new Multimodal Cross-Attention Fusion (MCAF) layer incorporating fine-grained token-region alignment through our proposed Semantic Graph Alignment Network (SGAN). An additional commonsense knowledge augmentation layer using ConceptNet is also integrated into our model, enabling eventual robust reasoning across open-domain settings. The extensive experiments on VQA v2, RefCOCO/+/g and MS-COCO captioning benchmarks present that MGLAN achieves the state-of-the-art results with an overall accuracy of 83.74% for VQA v2( surpassing its best previous model UNIMO by 4.44 percentage points) and 81.63% for RefCOCO visual grounding, while keeping a lighter parameter footprint (118M), along with ablation studies to verify each proposed component. We push the frontier of multimodal language grounding, and provide a scalable framework for real world applications such as scene understanding, visual question answering and multimodal dialogue systems.
关键词
Multimodal language processing, visual grounding, cross-modal attention, visual question answering, scene understanding, transformer, knowledge augmentation, NLP, deep learning, Vision Transformer
报告人
Elamathi E
ASSISTANT PROFESSOR Trichy；K.Ramakrishnan College of Engineering； Samayapuram稿件作者
全部评论
重要日期
-
会议日期
11月20日
2026
至11月22日
2026
-
10月31日 2026
初稿截稿日期
-
11月07日 2026
提前注册日期
-
11月22日 2026
注册截止日期
主办单位
United Societies of Science
承办单位
Pondicherry University
协办单位
IEEE Section
IEEE Madras Section
IEEE Madras Section
联系方式
- ICon INDIA 2026
- ic******@usssociety.org
发表评论