Sentiment analysis of video danmakus based on MIBE-RoBERTa-FF-BiLSTM

Area of Science:

• Natural Language Processing
• Affective Computing
• Multimedia Analysis

Background:

• Danmakus, or user-generated comments, offer real-time viewer interaction and sentiment insights crucial for video platforms.
• Traditional sentiment analysis struggles with danmaku's unique characteristics, including low transferability, segmentation issues, annotation inconsistency, and inadequate semantic feature extraction.

Purpose of the Study:

• To develop an advanced video danmaku sentiment analysis method addressing limitations of existing approaches.
• To improve the accuracy and robustness of sentiment classification for danmaku texts.
• To create a specialized dataset and domain lexicon for danmaku sentiment analysis.

Main Methods:

• A novel method, MIBE-RoBERTa-FF-BiLSTM, was proposed for video danmaku sentiment analysis.
• A "Bilibili Must-Watch List and Top Video Danmaku Sentiment Dataset" was constructed with 10,000 danmaku texts.
• A new word recognition algorithm using mutual information (MI) and branch entropy (BE) identified 2610 new words, forming a domain lexicon.
• Maslow's hierarchy of needs theory guided consistent sentiment annotation.
• The domain lexicon was integrated into the RoBERTa-FF-BiLSTM model for enhanced feature learning.

Main Results:

• The proposed MIBE-RoBERTa-FF-BiLSTM model achieved a superior F1 score of 94.06% on the danmaku sentiment classification task.
• Comparative experiments confirmed the model's best comprehensive performance, accuracy, and robustness over mainstream methods.
• The integrated domain lexicon effectively enhanced semantic feature extraction from danmaku texts.

Conclusions:

• The MIBE-RoBERTa-FF-BiLSTM method significantly advances video danmaku sentiment analysis.
• The study highlights the importance of domain-specific lexicons and advanced deep learning architectures for analyzing user-generated content.
• Future work should address manual lexicon construction and incorporate broader semantic information from video content.