ArguFormer

v2 in progress

ArgueFromer is a NLP system for analyzing debate transcripts using both traditional machine learning and modern transformer models, enabling comparative analysis of argumentative language.

The system detects:

• Speaker segmentation
• Sentiment of arguments
• Toxic language
• Logical fallacies
• Model comparison between classical NLP and transformer models

It demonstrates the evolution of NLP approaches from feature-engineered models to transformer-based language models.

---

Table of Contents

• Overview
• System Architecture
• Installation
• Project Structure

---

Overview

Debates often contain complex argumentative structures including emotional appeals, logical fallacies, and toxic rhetoric. Automatically analyzing such patterns can help in studying argument quality and discourse dynamics.

The Arguformer processes debate transcripts and performs multi-stage analysis:

1. Speaker Segmentation Extracts individual statements from debate transcripts.

2. Sentiment Analysis Detects whether arguments express positive or negative sentiment.

3. Toxicity Detection Identifies abusive or harmful language.

4. Fallacy Detection

   • Version 1 (Traditional NLP) Uses TF-IDF + Logistic Regression.
   • Version 2 (Transformer NLP) Uses DistilBERT fallacy classification.

5. Model Comparison Both models are executed in parallel and their predictions are compared to produce a final judgement.

This approach highlights the strengths and limitations of both classical NLP pipelines and transformer models.

---

System Architecture

Debate Transcript
        │
        ▼
Speaker Segmentation
        │
        ▼
Sentence Level Analysis
        │
 ┌───────────────┬─────────────────┐
 │               │                 │
 ▼               ▼                 ▼
Traditional NLP  Sentiment Model   Toxicity Model
TF-IDF + LR      (DistilBERT)       (Toxic-BERT)
 │
 ▼
Fallacy Prediction
 │
 ▼
Transformer Fallacy Model
(DistilBERT)
 │
 ▼
Model Comparison Engine
 │
 ▼
Final Debate Analysis Output

The architecture enables parallel analysis pipelines, allowing comparison between feature-engineered ML methods and deep transformer models.

---

Installation

1. Clone the Repository

git clone https://github.com/Atim45/ArguFormer.git
cd ArguFormer

2. Install Dependencies

pip install -r requirements.txt

3. Run the Project

python main.py

Example input debate:

Speaker A: You are the dumbest person alive
Speaker B: That argument makes no sense
Speaker A: Think about the poor children suffering

Example output:

Speaker   Sentence                         Sentiment   Toxicity   Final Judgement
----------------------------------------------------------------------------------
Speaker A You are the dumbest person alive NEGATIVE    toxic      Ad Hominem
Speaker B That argument makes no sense     NEGATIVE    toxic      Strawman
Speaker A Think about the poor children    POSITIVE    non-toxic  Appeal to Emotion

---

Project Structure

arguformer/

├── cli/
│   ├── __init__.py
│   └── args.py

├── core/
│   ├── __init__.py
│   ├── preprocessing.py
│   ├── segmentation.py
│   └── comparison_engine.py

├── models/
│   ├── __init__.py
│   ├── fallacy_lr.py
│   ├── fallacy_transformer.py
│   └── manager.py   ← soon

├── utils/
│   ├── __init__.py
│   ├── logger.py
│   ├── exceptions.py
│   └── config_loader.py

├── configs/
│   └── default.yaml

├── datasets/
├── evaluation/
├── logs/
├── outputs/
├── tests/
├── rag/

├── main.py
├── requirements.txt
├── README.md
└── .gitignore
---

# Technologies Used

* Python
* HuggingFace Transformers
* Scikit-Learn
* TF-IDF Vectorization
* Logistic Regression
* DistilBERT
* Toxic-BERT

---

# Future Improvements

Potential future upgrades include:

* Larger **debate datasets**
* Improved **fallacy classification models**
* **Visualization dashboards**
* **Argument strength scoring**
* **persuasion technique detection**