youtube-travel-buzz-sentiment-classifier

youtube-travel-buzz-sentiment-classifier

🔹 Model Name

youtube-travel-buzz-sentiment-classifier

---

🔹 Model Description

A Korean multi-class sentiment classifier that decomposes travel-related YouTube comments into positive, neutral, and negative signals for travel demand analysis.

---

🔹 Model Summary

This model performs three-class sentiment classification on Korean YouTube comments that have already been identified as travel-related.

Sentiment labels:

• 0: Negative
• 1: Neutral
• 2: Positive

Unlike conventional sentiment models, this classifier explicitly preserves neutral sentiment, which primarily captures information-seeking and intent-driven comments.

This design enables downstream analysis linking online discourse patterns to real-world travel demand signals.

The model is trained on LLM-generated synthetic comments designed to mimic the linguistic characteristics of real YouTube travel discussions.

---

🔹 Intended Use

Primary Use Case

• Decomposing travel-related YouTube buzz into structured sentiment signals
• Supporting:
  • Exploratory demand analysis
  • Early-stage travel interest detection
  • Trend-level behavioral research

Out-of-Scope Use

• Emotion detection beyond sentiment polarity
• Individual-level behavior prediction
• Standalone decision-making systems

---

🔹 Training Data

• Type: Synthetic Korean YouTube travel comments generated using multiple LLMs
• Labels:
  • 0: Negative
  • 1: Neutral
  • 2: Positive
• Key Characteristics:
  • Informal language, slang, typos, emojis
  • Mixed sentence length and ambiguity
  • Designed to approximate real-world YouTube comment noise
  • Neutral comments intentionally modeled to represent questions, factual statements, and information-seeking behavior

Downstream analysis revealed that neutral sentiment often functions as a proxy for latent travel intent, particularly for emerging destinations.

Prompt design details and data generation strategy are documented in the associated GitHub repository.

---

🔹 Model Architecture

• Base model: monologg/koelectra-base-discriminator
• Task: Multi-class sentiment classification
• Tokenizer: KoELECTRA tokenizer
• Fine-tuning: Hugging Face Trainer API

---

🔹 Performance (Indicative)

• Overall Accuracy: ~96%
• Macro F1-score: ~96% (balanced synthetic validation set)

These metrics were obtained on a held-out synthetic validation set and reflect controlled experimental conditions.

Given the semantic ambiguity of short-form YouTube comments, the model is intended for trend-level and aggregate analysis rather than individual comment-level judgment.

Performance on real-world YouTube comments may differ due to distribution shift and unmodeled linguistic nuance.

---

🔹 Limitations

• Fine-grained emotional nuance is not modeled

• Synthetic data bias may persist in edge cases

• Not optimized for sarcasm-heavy or long-form comments

• Performance may degrade on real-world comments without additional fine-tuning on authentic data.

---

🔹 Ethical Considerations

• No personal data used
• Outputs should be interpreted at aggregate signal level, not individual judgment

---

🔹 Related Resources

• 📁 Full pipeline code and documentation:

  https://github.com/DalDream/youtube-travel-buzz-nlp-pipeline

• 🔗 Upstream travel relevance classifier:

  https://huggingface.co/DalDream/youtube-travel-buzz-relevance-classifier

---

🔹 Citation / Attribution

This model was developed as part of a YouTube Travel Buzz Signal Extraction NLP pipeline for research and portfolio demonstration purposes.

Author / Contributions

• [DalDream] – Project lead for model strategy, pipeline design, model validation, and final documentation.
• [GY Yu] – LLM-based synthetic data generation, dataset construction, model training, and fine-tuning.

Note: This model is the result of a collaborative team project. Responsibilities are listed to clarify individual contributions.