The exponential growth of information available online has created a challenge for individuals to effectively search and process the vast amount of data. Automatic text summarization emerged as a solution that can enhance information accessibility by providing a concise overview. Recent advancements in large pre-trained language models achieve remarkable performance on many summarization tasks. However, these models still face limitations when dealing with non-regular input types. In this thesis, we present two different types of summarization tasks. In the first half, we focus on meeting decision summarization. To achieve this, we propose a method that takes the entire meeting transcript as input. This method involves identifying salient words, finding sentences related to decisions, clustering decision-related sentences, and then generating summaries for each cluster. Our approach surpasses previous methods for this meeting summarization subtask. In the second half, we tackle the task of video captioning. We introduce a two-step method that first generates captions for frames within the video, followed by summarizing these generated captions to form concise video captions. Our experimental results demonstrate that our proposed method outperforms previous video-language generation models, highlighting its enhanced performance using summarization.