PakDramaEcho: A DistilBERT-Labeled Urdu Sentiment Dataset

Description

📊 PakDramaEcho Dataset Description & Research Interpretation 🧠 Research Hypothesis This study assumes that sentiment expressed in YouTube comments on Pakistani drama content can be effectively modeled using transformer-based NLP (DistilBERT). It further hypothesizes that these sentiments reflect meaningful emotional patterns linked to storylines, characters, and dramatic events in Urdu dramas. Specifically: * Informal Urdu/Roman Urdu still contains strong sentiment signals despite noise * DistilBERT can perform reliable sentiment labeling in low-resource settings * Viewer comments reflect real audience emotional response to dramas 📦 Data Overview PakDramaEcho is a sentiment analysis dataset created from YouTube comments on Pakistani drama videos. Source: YouTube drama comment sections Language: Urdu, Roman Urdu, mixed English-Urdu Domain: Pakistani TV dramas Task: Sentiment classification (Positive / Neutral / Negative) Labeling Method: DistilBERT-based automatic annotation Format: CSV file 🧹 Data Collection & Processing Collection * Scraped publicly available YouTube comments from drama-related videos * Only public comments were included Preprocessing * Removed URLs, emojis, and special characters * Cleaned repeated/noisy text * Normalized Urdu text * Filtered empty/irrelevant comments * Optional deduplication applied Labeling * Sentiment labels generated using DistilBERT classifier * Classes: Positive, Neutral, Negative 📈 Key Insights 1. Strong Emotional Engagement Most comments are emotionally expressive, especially positive ones, showing strong audience connection with drama content. 2. Noisy & Informal Language Includes Roman Urdu, spelling variations, and mixed-language text, reflecting real-world social media challenges. 3. Sentiment Imbalance Positive sentiment dominates, likely due to fan engagement and selective commenting behavior. 4. Context-Dependent Emotion Sentiment depends on characters, emotional scenes, and plot twists rather than isolated words. 🔍 Interpretation This dataset should be treated as a real-world noisy corpus rather than a clean benchmark. It represents: * Audience perception, not objective truth * Emotional response to entertainment media * A resource for low-resource Urdu NLP research ⚠️ Since labels are generated using DistilBERT, there may be classification noise and model bias. 🚀 Potential Use Cases * Urdu sentiment classification benchmarking * Transformer model training (BERT, RoBERTa, DistilBERT) * Low-resource NLP and domain adaptation * Social media opinion mining * Audience behavior analysis ⚠️ Limitations * Auto-generated labels (not human verified) * Domain-specific (Pakistani dramas only) * Noisy informal text * Possible sentiment imbalance

Steps to reproduce

⚙️ Data Acquisition and Reproducibility Protocol (PakDramaEcho) 1. Data Source The dataset was collected from publicly available YouTube comments on Pakistani drama videos selected based on popularity and relevance. Only public user comments were used. 2. Data Collection Comments were extracted using the YouTube Data API v3 or scraping scripts (Python-based). Video URLs were manually curated, and comment threads were automatically extracted and stored in CSV format using pandas. 3. Tools Used Python, YouTube API / scraping tools (e.g., Selenium if needed), pandas, and standard data processing libraries. 4. Preprocessing Raw comments were cleaned by removing URLs, emojis, duplicates, and noise. Text was normalized through lowercasing, whitespace removal, and basic script standardization. Spam and irrelevant entries were filtered out. 5. Sentiment Labeling Labels (Positive, Neutral, Negative) were generated using a pretrained DistilBERT sentiment classifier. This automated labeling was used as pseudo-annotation for scalability. 6. Dataset Structure Final dataset contains: text: cleaned comment label: sentiment class (Optional: video_id, timestamp) 7. Reproducibility Steps Select Pakistani drama videos on YouTube Extract comments via API or scraper Clean and preprocess text Apply DistilBERT sentiment model Save as structured CSV 8. Environment Python ≥3.8, pandas, numpy, transformers, torch/tensorflow, scikit-learn. 9. Ethics Only public data was used, with anonymization applied. The dataset is intended for academic research only. Workflow Summary Video selection → Comment extraction → Cleaning → Preprocessing → DistilBERT labeling → CSV export

Institutions

• Port City International University

  Chittagong, Chittagong

Categories

Licence