<p>Manually gaging public sentiment in the millions of informal, multilingual comments posted on social media vlogs is infeasible, and existing models often falter on slang, code-switching, emojis, and rapid shifts in tone. To address this challenge, we compile a 36 000-comment corpus drawn from India’s most-viewed channels on YouTube, Twitter (X), TikTok, Reddit, Instagram, and Facebook, then build an end-to-end Python NLP pipeline that blends deep learning–based preprocessing (tokenization, misspelling correction, code-mix detection) with an enhanced version of VADER (Valence Aware Dictionary for Sentiment Reasoning). The lexicon is augmented with 276 domain-specific slang terms and emojis, and custom rules capture sarcasm, negation, and intensity cues. The upgraded VADER attains an F1-score of 1.00 on the in-sample corpus and 0.92 on an external hold-out set, surpassing fine-tuned BERT (F1 = 0.94) and zero-shot GPT<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="41060_2025_849_Article_IEq1.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$-$</EquationSource> <EquationSource Format="MATHML"><math> <mo>-</mo> </math></EquationSource> </InlineEquation>3.5 (F1 = 0.91). Regression metrics (MAE = 1.4, MSE = 3.1, <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="41060_2025_849_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="71" /> </InlineMediaObject> <EquationSource Format="TEX">$R^{2}=0.99$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.99</mn> </mrow> </math></EquationSource> </InlineEquation>) confirm precise sentiment intensity estimation, while temporal analysis reveals highly polarized reactions within the first 24 h that converge toward neutrality over time. Content-type analysis shows strong positivity for technology and fashion vlogs (51–57%) but 61% disapproval for travel content. Overall, the framework generalizes across comedic, scientific, and satirical contexts (36.8% positive, 57.3% neutral) and offers a transparent, scalable solution for real-time sentiment monitoring in multilingual social media environments.</p>

Video sentiment analysis on social media using an advanced VADER technique

期刊论文
发表日期：

2025年7月23日

查看全文

Muskan Dixit,
Malvinder Singh Bali,
Kanwalpreet Kour,
Iacovos Ioannou,
G. S. Pradeep Ghantasala,
Vasos Vassiliou

摘要

Manually gaging public sentiment in the millions of informal, multilingual comments posted on social media vlogs is infeasible, and existing models often falter on slang, code-switching, emojis, and rapid shifts in tone. To address this challenge, we compile a 36 000-comment corpus drawn from India’s most-viewed channels on YouTube, Twitter (X), TikTok, Reddit, Instagram, and Facebook, then build an end-to-end Python NLP pipeline that blends deep learning–based preprocessing (tokenization, misspelling correction, code-mix detection) with an enhanced version of VADER (Valence Aware Dictionary for Sentiment Reasoning). The lexicon is augmented with 276 domain-specific slang terms and emojis, and custom rules capture sarcasm, negation, and intensity cues. The upgraded VADER attains an F1-score of 1.00 on the in-sample corpus and 0.92 on an external hold-out set, surpassing fine-tuned BERT (F1 = 0.94) and zero-shot GPT $-$ $-$ 3.5 (F1 = 0.91). Regression metrics (MAE = 1.4, MSE = 3.1, $R^{2}=0.99$ $R^{2} = 0.99$ ) confirm precise sentiment intensity estimation, while temporal analysis reveals highly polarized reactions within the first 24 h that converge toward neutrality over time. Content-type analysis shows strong positivity for technology and fashion vlogs (51–57%) but 61% disapproval for travel content. Overall, the framework generalizes across comedic, scientific, and satirical contexts (36.8% positive, 57.3% neutral) and offers a transparent, scalable solution for real-time sentiment monitoring in multilingual social media environments.