MIBE RoBERTa FF BiLSTM: A Hybrid Deep Learning Framework for Sentiment Analysis of Video Danmakus
Abstract
The complexity of visual information combined with the subjective character of human emotions makes emotion recognition from visual data videos a formidable obstacle. Among the many computer vision tasks that deep learning has proven adept at over the years is sentiment categorization. This research presents a video danmaku sentiment-analysis (SA) approach based on MIBE-RoBERTa-FF-BiLSTM to address the issues of poorly transferable classical SA approaches to the danmaku domain, inaccurate danmaku text segmentation, inconsistent sentiment explanation, and inadequate semantic feature extraction. Based on our own research, this article compiles a "Bilibili Must-Watch List and Top Video Danmaku Sentiment Dataset" that includes 10,000 danmaku texts spanning 18 different themes, both positive and negative. In order to create a domain lexicon, a novel word recognition technique that utilizes branch entropy (BE) and mutual information (MI) is employed to unearth 26,10 popular new words in the dataset that are irregular networks, ranging from trigrams to heptagrams. For sentiment classification in danmaku texts, the RoBERTa-FF-BiLSTM model incorporates the domain lexicon into its feature fusion layer, allowing it to learn all of the semantic properties of words, characters, and contexts. The proposed model in this study outperforms existing existing techniques for video danmaku text SA in terms of comprehensive performance, accuracy, and resilience, according to experimental results on the dataset. The model’s F1 value is 98.06%.