Executive Summary : | This project proposal aims to design frameworks for federated meta-learning to solve real-life natural language processing problems. Meta-learning-based approaches have become popular for enhancing the generalization capability of machine learning-based models. As different agencies struggle to share data due to confidentiality issues, federated learning-based frameworks have become a popular choice. The project aims to design generalized machine learning models using meta-learning concepts in the federated learning framework. One real-life application is multi-modal summarization (MMS), which involves generating summaries using different modalities like text, audio, and video. The increasing amount of information on the internet makes it difficult for users to comprehend useful information, necessitating research on the task of MMS. Various studies have shown that including multi-modal data as input can improve summary quality. The MMS task is demanding and challenging, as humans generating a multi-modal summary must use their prior understanding and external knowledge to produce an abstract. Establishing computer systems to mimic this behavior becomes onerous due to their lack of human perception and knowledge. Multimodal summarization has various applications, such as meeting minutes summarization, video summarization, automatic content creation, sports summarization, and news summarization. In cases where agencies have private data that they are not ready to share due to privacy issues, federated learning-based approaches can be a practical solution. The limited amount of data available with agencies and the varying nature of multi-modal information make meta-learning essential in developing generalized models. Therefore, federated meta-learning-based frameworks are required to solve the task of multi-modal summarization. |