Scholarlink

<p>Aiming to address the technical bottleneck of difficult deployment and poor real-time performance in traditional deep learning models for sentiment analysis of folk culture short videos, this study proposes a lightweight BiLSTM and edge computing integration model. Through model pruning and quantitative compression technology, the number of parameters is reduced by 80.5%, while achieving an accuracy rate of 88.5% in the folk culture emotion classification task. The F1 value reaches 87.9%, which is significantly higher than that of the traditional model. The model integrates an exclusive folk culture thesaurus and an attention mechanism to strengthen cultural semantic learning; ablation results indicate that the knowledge-guided weight <i>w</i><sub><i>k</i></sub> derived from this lexicon improves the F1-score by 4.2% compared to a baseline without the thesaurus. Experimental verification: The lightweight BiLSTM has a parameter volume of 2.1&#xa0;M and a standalone model inference speed of 45ms per video, while the integrated system achieves an end-to-end delay ranging from 220 to 350ms, which is significantly better than the BERT-base and standard CNN benchmarks. In edge computing scenarios, the average delay of 5-second video clips is 220ms, the delay of 8-second clips is 280ms, and the delay of 75% samples is less than 300ms. The delay is stable at 250ms when there are 10 concurrent requests, meeting real-time requirements. In terms of power consumption, the high-quantization accuracy model consumes about 5000mW at 100FPS, and the medium-quantization model reduces power to 4000mW at the same throughput, which is better than the 4500mW of the benchmark unquantized model. Fine-grained analysis shows that the accuracy rate of positive emotions in traditional festivals is 93.2%, that of folk dances is 91.5%, that of local operas is 89.8% and that of handicrafts is 88.0%. This study presents a high-performance and low-power solution for sentiment analysis of folk culture content, offering both theoretical significance and practical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on sentiment analysis model of folk culture short videos integrating lightweight BiLSTM and edge computing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Ziqing Lin

摘要

Aiming to address the technical bottleneck of difficult deployment and poor real-time performance in traditional deep learning models for sentiment analysis of folk culture short videos, this study proposes a lightweight BiLSTM and edge computing integration model. Through model pruning and quantitative compression technology, the number of parameters is reduced by 80.5%, while achieving an accuracy rate of 88.5% in the folk culture emotion classification task. The F1 value reaches 87.9%, which is significantly higher than that of the traditional model. The model integrates an exclusive folk culture thesaurus and an attention mechanism to strengthen cultural semantic learning; ablation results indicate that the knowledge-guided weight wk derived from this lexicon improves the F1-score by 4.2% compared to a baseline without the thesaurus. Experimental verification: The lightweight BiLSTM has a parameter volume of 2.1 M and a standalone model inference speed of 45ms per video, while the integrated system achieves an end-to-end delay ranging from 220 to 350ms, which is significantly better than the BERT-base and standard CNN benchmarks. In edge computing scenarios, the average delay of 5-second video clips is 220ms, the delay of 8-second clips is 280ms, and the delay of 75% samples is less than 300ms. The delay is stable at 250ms when there are 10 concurrent requests, meeting real-time requirements. In terms of power consumption, the high-quantization accuracy model consumes about 5000mW at 100FPS, and the medium-quantization model reduces power to 4000mW at the same throughput, which is better than the 4500mW of the benchmark unquantized model. Fine-grained analysis shows that the accuracy rate of positive emotions in traditional festivals is 93.2%, that of folk dances is 91.5%, that of local operas is 89.8% and that of handicrafts is 88.0%. This study presents a high-performance and low-power solution for sentiment analysis of folk culture content, offering both theoretical significance and practical applications.