🌳 EcoDigest

🧠 Sentiment & Consensus Analyzer

A multi-stage NLP pipeline designed to analyze environmental sentiment and extract representative consensus from public comments. The project leverages fine-tuned transformer models for classification and summarization, combined with semantic clustering for representative insight extraction.

---

🚀 Applications

1. Unified Sentiment & Consensus Analyzer

File: sentiment_summarizer_app.py
The primary interactive tool. Input a list of raw comments and the AI will:

1. Sentiment Analysis: Classify each comment as Positive or Negative using a fine-tuned RoBERTa model.
2. Centroid Ranking: Identified the most representative "centroid" comments for each group using MiniLM embeddings.
3. Consensus Summarization: Generate a concise summary of the group's collective opinion using a fine-tuned FLAN-T5 model.

To Run:

streamlit run sentiment_summarizer_app.py

2. Public Perception Dashboard

File: dashboard_app.py
A visualization suite for large-scale batch analysis (requires model_output.csv).

• Sentiment distribution and over-time trends.
• Interactive Word Clouds and composición analysis.
• Raw data explorer.

To Run:

streamlit run dashboard_app.py

---

🛠 Project Structure

• Model_ROBERTA-Sentiment/: Fine-tuned model for environmental sentiment classification.
• Model_FLAN-T5/: Fine-tuned model for high-fidelity summarization.
• utils.py: Centralized logic for model loading and inference pipelines.
• datasets/: Training and testing data, including the summarization_dataset.csv for sample testing.
• tests/: Standalone CLI scripts for testing individual model performance (test_sentiment.py, test_summ.py).
• requirements.txt: Full dependency list.

---

🧬 Scientific Pipeline

1. Sentiment Classification: Based on cardiffnlp/twitter-roberta-base-sentiment, fine-tuned on climate-specific datasets.
2. Ranking (Centroid): Uses all-MiniLM-L6-v2 to vectorize comments. We calculate the mean vector (centroid) of each cluster and select the top $k$ comments with the highest cosine similarity to the centroid.
3. Summarization: A sequence-to-sequence model (FLAN-T5) fine-tuned with specific prompts:
   • Positive: "Summarize the opinions of users who believe in the reality of climate change.:"
   • Negative: "Summarize the opinions of users who skeptical of/deny climate change.:"

---

📋 Installation

1. Create and activate a environment (Conda or venv):

conda create -n sentiment-analysis python=3.12
conda activate sentiment-analysis

2. Install dependencies:

pip install -r requirements.txt

3. Ensure models are placed in the root directory (refer to the Model Files section in structure).

---

📚 Original Research & Training

The core models were developed and analyzed in the following notebooks:

• Model_Training.ipynb: Original RoBERTa fine-tuning process.
• FLANT5-FT.ipynb: Dataset preparation and fine-tuning for the summarization model.
• Advanced_Sentiment_Analytics.ipynb: Initial exploration of keyword extraction and NER.

🏢Team Members

• Lawrence Agarin (lawrenceivanpagarin@iskolarngbayan.pup.edu.ph)
• Kyle Desmond Co (kyledesmondpco@iskolarngbayan.pup.edu.ph)
• Reymel Sardenia (reymelosardenia@iskolarngbayan.pup.edu.ph)
• Ken Satorre (kencalvinssatorre@iskolarngbayan.pup.edu.ph)
• Earl Andrei Fidel (earlandreidfidel@iskolarngbayan.pup.edu.ph)

🖥️ Bachelor of Science in Computer Science

🏫 Special Thanks

Polytechnic University of the Philippines - Manila
Ria A. Sagum - Instructor (Natural Language Processing)

🙏 Acknowledgements

This project makes use of several open-source libraries and pre-trained models. We would like to acknowledge the contributions of the following:

• Hugging Face: For the transformers library and hosting the model hub.
• Cardiff NLP: For the base twitter-roberta-base-sentiment model.
• Google Research: For the FLAN-T5 architecture.
• Sentence Transformers: For the all-MiniLM-L6-v2 model used in semantic ranking.
• Streamlit: For the interactive web application framework.
• Scientific Python Community: Including the developers of PyTorch, Pandas, NumPy, and Scikit-learn.
• Public Sentiment Analysis on Climate Change (simran98solanki): for the Twitter Sentiment Dataset.