VidLyze: Video Content Toxicity Identification Tool Using Natural Language Processing (NLP) and Machine Learning

VidLyze: Video Content Toxicity Identification Tool Using Natural Language Processing (NLP) and Machine Learning

연구 분야: Artificial Intelligence

학회: International Symposium on Intelligent Computing Systems

Toxicity identification and classification in multimedia contents is becoming an important field of research, particularly with the massive amount of multimedia contents that are produced and published on the web. Toxic multimedia content such as the use of inappropriate and bad words and swearing are not suitable for general audiences, especially for young children who are highly impressionable. In this paper, we present our current work-in-progress on designing and developing analytics model known as VidLyze for identifying different levels of toxicity in a given video, based on the text analysis approach. The goal of VidLyze is to use machine learning and natural language processing (NLP) to make a system that can watch and analyze video material for toxic contents. The goal is to find toxic, inappropriate or insulting content in videos so that people can watch them safely. The work carried out involved extracting video contents with different formats, converting audio to text in order to obtain the audio transcript, and performing text toxicity analysis on the audio transcript using a suitable machine learning algorithm. The proposed analytical model requires NLP integration for analyzing human languages from video and audio, and the need for datasets that are large enough to train the respective model. We adopt the Long Short-Term Memory (LSTM) recurrent neural networks to train the model for classifying toxicity based on large textual datasets. VidLyze is made to be a safe and reliable method that makes content moderation better and makes sure everyone has a good time watching.