🌍 ClimateNet: A Multi-Faceted Classifier for Analyzing Climate Change Discourse

Authors: ^$Tustee Mazumdar, ^$Adrika Chowdhury, Abu Nowshed Chy, and Tarem Ahmed || $- Equal Contribution ||

---

Publication Status

This paper, “ClimateNet: A Multi-Faceted Classifier for Analyzing Climate Change Discourse,” has been accepted and presented at the IEEE Region 10 Humanitarian Technology Conference (R10-HTC 2025).
It is currently in press and will be published soon in the IEEE Xplore digital library.

---

Overview

ClimateNet is a multi-task text classification framework designed to analyze the multifaceted discourse surrounding climate change on social media. It leverages transformer-based contextual embeddings (BERT, RoBERTa, DistilBERT, ClimateBERT) integrated with a Bidirectional LSTM (BiLSTM) and a Multilayer Perceptron (MLP) classifier to capture both semantic and sequential dependencies in textual data.

The model aims to uncover diverse perspectives—including stance, relevance, humor, and hate—within climate-related conversations, aligning with UN Sustainable Development Goal (SDG) 13: Climate Action.

---

⚙️ Architecture

The proposed ClimateNet framework follows a hybrid multi-model architecture that combines transformer-based feature extraction with sequential and dense learning modules for robust text classification.

Each transformer (BERT, RoBERTa, DistilBERT, ClimateBERT) is fine-tuned and followed by a BiLSTM and MLP classifier.

• Transformers: Extract deep contextual features
• BiLSTM: Captures long- and short-term dependencies
• MLP: Maps learned representations to output labels

---

In essence, ClimateNet integrates contextual transformers with temporal sequence modeling and dense classification layers, enabling the framework to accurately predict multi-faceted aspects of climate change discourse across six tasks.

---

Tasks

ClimateNet performs six independent NLP classification tasks on the ClimaConvo dataset:

Task	Description	Label Type
Relevance Detection	Classify whether a tweet is related to climate change	Binary (0–Non-relevant, 1–Relevant)
Stance Detection	Identify stance: Support, Oppose, or Neutral	Multiclass (0–Support, 1–Oppose, 2–Neutral)
Hate Speech Detection	Detect presence of hate content	Binary (0–Non-hate, 1–Hate)
Direction of Hate Speech	Determine if hate is direct or indirect	Binary (0–Undirected, 1–Directed)
Target of Hate Speech	Identify if hate targets individuals, organizations, or communities	Multiclass (0–Individual, 1–Organization, 2–Community)
Humor Detection	Detect humorous or sarcastic tone	Binary (0–Non-humor, 1–Humor)

---

Dataset

We used ClimaConvo, a publicly available dataset comprising 15,309 English climate-related tweets, annotated across the six dimensions listed above.
Data split: 80% train / 20% test (stratified).
Main evaluation metric: Macro-averaged F1-score.

---

📂 Repository Structure

Notebooks/
├── Relevance/
│   ├── BERT_Relevance.ipynb
│   ├── ClimateBERT_Relevance.py
│   ├── DistilBERT_Relevance.py
│   └── RoBERTa_Relevance.py
│
├── Stance/
│   ├── BERT_Stance.ipynb
│   ├── ClimateBERT_Stance.ipynb
│   ├── DistilBERT_Stance.ipynb
│   └── RoBERTa_Stance.ipynb
│
├── Hate_Speech/
│   ├── BERT_HateSpeech.ipynb
│   ├── ClimateBERT_HateSpeech.ipynb
│   ├── DistilBERT_HateSpeech.ipynb
│   └── RoBERTa_HateSpeech.ipynb
│
├── Hate_Speech_Directions/
│   ├── BERT_Directions.ipynb
│   ├── ClimateBERT_Directions.ipynb
│   ├── DistilBERT_Directions.ipynb
│   └── RoBERTa_Directions.ipynb
│
├── Hate_Speech_Targets/
│   ├── BERT_Targets.ipynb
│   ├── ClimateBERT_Targets.ipynb
│   ├── DistilBERT_Targets.ipynb
│   └── RoBERTa_Targets.ipynb
│
└── Humor/
    ├── BERT_Humor.ipynb
    ├── ClimateBERT_Humor.ipynb
    ├── DistilBERT_Humor.ipynb
    └── RoBERTa_Humor.ipynb

---

Inside the Notebooks directory, each subdirectories correspond to a specific classification task, and within each folder, four transformer-based implementations are provided separately for modularity and reproducibility. All scripts include:

• Dataset preprocessing and tokenization
• Model fine-tuning and BiLSTM integration
• Training, evaluation, and result saving

---

Installation

To replicate our experiments or explore the ClimateNet notebooks locally, please follow these steps:

Clone the Repository

git clone https://github.com/<your-username>/ClimateNet.git
cd ClimateNet

Set Up the Python Environment

Use Python 3.10+

Install Dependencies

pip install -r requirements.txt

Run the Notebooks

Upload the notebook of your choice (e.g., Notebooks/Stance/DistilBERT_Stance.ipynb) to Colab and ensure the runtime is set to GPU for faster training.

Acknowledgment

This research builds upon the ClimaConvo dataset introduced by Shiwakoti et al. (2024) in their paper “Analyzing the Dynamics of Climate Change Discourse on Twitter: A New Annotated Corpus and Multi-Aspect Classification” (LREC-COLING 2024).
We gratefully acknowledge their contribution for providing the annotated benchmark dataset that enabled our experiments and evaluation.