Context-aware video caption generation with VideoBERT and domain-aware adapters

Abstract

Background. Automated video description is a relevant task for workplace reengineering and business process analysis, yet its  real-world efficiency is affected by domain shift between training datasets and deployment environments. Aim. The study aims to  develop a context-aware video caption generation architecture combining pretrained video representations with lightweight  adaptation mechanisms. Methods. The proposed framework integrates a frozen twelve-layer VideoBERT encoder with a dual-branch  Domain-Aware Adapter adaptation module and a three-layer autoregressive Transformer decoder. Visual features are extracted using  an ImageNet-pretrained ResNet-50 backbone. The Domain-Invariant Adapter and Domain-Specific Adapter branches process  contextualized representations in parallel and are combined through a trainable scalar fusion coefficient. A fifty percent stochastic  Masked Language Modeling strategy is applied during training. Results. Evaluation on the Something-Something V2 benchmark  shows that the complete VideoBERT, Domain-Aware Adapter and Transformer decoder configuration with fifty percent Masked  Language Modeling achieves a Mean BLEU-4 of zero point one four six nine, Mean ROUGE-L of zero point three four seven four,  and Mean METEOR of zero point two seven two zero. Component-wise ablation demonstrates improvements from both the dual branch adaptation module and Masked Language Modeling regularization. Conclusions. The proposed architecture provides a  modular, parameter-efficient framework for context-aware video caption generation, while explicit cross-domain transfer evaluation  remains a direction for future research.