The IEEE Workshop on Parallel and Distributed Processing for Computational Social Systems (ParSocial) serves as a platform bringing together researchers from interdisciplinary areas of computer science, social sciences, applied mathematics, and engineering to further research on parallel and distributed processing, computational modeling, and high-performance computing in the discipline of social computing.
At the end of May 2024, the 8th ParSocial conference will be held in San Francisco, California. There Mert Can Cakmak and Dr. Nitin Agarwal will present their research titled “High-Speed Transcript Collection on Multimedia Platforms: Advancing Social Media Research through Parallel Processing.” This research uses parallel processing to improve the processing time of retrieval of video captions, video transcripts, and transcript translations. They also achieved a faster processing time by using an AI model specially created for faster speeds, and by customizing the different parameters of the AI model that was used for transcript generation.
To identify how effective the parallelization was, they compared an initial benchmark (where transcripts, translations, and captions were retrieved with no changes to the retrieval method) to the times after their changes. By spreading out each task across 8 GPUs, they were able to achieve much faster times. This, combined with an OpenAI Whisper model designed for speed and with parallelization, made transcript generation 90 times faster. Similarly, utilizing parallelization and multiple GPUs, translation was increased by an amazing 36,905 times. Likewise, caption retrieval improved by 22 times when 8 GPUS were used to parallelize the task. These results prove that parallelization significantly improves processing time of such tasks.
Dr. Agarwal said, “At COSMOS, we are constantly trying innovative ways to improve social computing techniques. To cope with the high volume and velocity of social media data, we have developed a distributed computing technique to process YouTube data. We are thrilled to have our study published at ParSocial 2024!”