A Computational Approach to Modeling Conversational Systems

Analyzing Large-Scale Quasi-Patterned Dialogue Flows

Official Implementation of the IEEE EUROCON 2025 Paper
A Computational Approach to Modeling Conversational Systems Analyzing Large-Scale Quasi-Patterned Dialogue Flows
Mohamed Achref Ben Ammar – National Institute of Applied Science and Technology (INSAT), University of Carthage, Tunisia
Mohamed Taha Bennani – University of Tunis El Manar (FST)

---

Abstract

The rise of large language models (LLMs) has led to increasingly complex and loosely structured dialogues. In this work, we introduce a computational graph-based framework that models these quasi-patterned conversations. Central to our approach is the Filter & Reconnect method, a graph simplification technique that reduces conversational noise while preserving semantic structure.

Key outcomes:

• 2.06× improvement in semantic metric S over prior methods
• 0 δ-hyperbolicity, enforcing a tree-like, interpretable structure

This framework offers practical tools for monitoring and analyzing chatbot behavior, dialogue management systems, and user interaction patterns at scale.

---

Methodology Overview

The methodology consists of the following core steps:

1. Utterance Extraction
   Conversational utterances are extracted from a structured dataset consisting of multi-turn dialogues.

2. Semantic Embedding
   Each utterance is transformed into a dense vector using a pre-trained text embedding model, capturing the semantic meaning of the message.

3. Clustering of Intents
   Using hierarchical clustering techniques and a large language model (LLM), similar utterances are grouped together to identify key communicative intents.

4. Markov Chain Construction
   A Markov Chain is built where nodes represent clustered intents and edges represent transitions between them in the dialogue flow.

5. Graph Simplification: Filter & Reconnect
   The conversational graph undergoes a noise reduction process by removing irrelevant transitions while preserving semantic and structural coherence.

6. Flow Pattern Analysis
   The resulting graph is then analyzed to identify quasi-patterned conversational flows, enabling improved interpretability and dialogue system evaluation.

---

Setup

1. Install Dependencies

# Create and activate a virtual environment
python -m venv venv
source venv/bin/activate        # Linux/MacOS
venv\Scripts\activate           # Windows

# Install required packages
pip install -r requirements.txt

# Download required NLP model
python -m spacy download en_core_web_md

2. Create a .env File

At the project root, create a .env file and configure the following environment variables:

# Python setup
PYTHONPATH=${PYTHONPATH}:.

# Environment mode
ENVIRONMENT="local"

# API Keys
GOOGLE_API_KEY=
MISTRAL_API_KEY=

Ensure your API keys are valid and have the appropriate access privileges.

---

Input Data Format

This framework supports ABCD v1.1, MultiWOZ 2.0, or any custom dataset formatted as follows:

{
  "conversation_1": [
    {"role": "agent", "content": "Hello, how can I help you today?"},
    {"role": "customer", "content": "I need assistance with my account."},
    {"role": "action", "content": "Agent opened account details."}
  ]
}

• Save your data file as: data/processed_formatted_conversations.json

---

Run the Pipeline

python main.py \
    --file_path data/processed_formatted_conversations.json \
    --num_sampled_data 500 \
    --min_clusters 10 \
    --max_clusters 30 \
    --model_name 'sentence-transformers/all-mpnet-base-v2' \
    --label_model 'open-mixtral-8x22b' \
    --tau 0.15 \
    --top_k 2 \
    --alpha 0.8

---

Advanced Configuration

Parameter	Description	Default
--num_sampled_data	Number of conversations to sample	100
--min_clusters	Minimum cluster count for elbow method	5
--max_clusters	Maximum cluster count for elbow method	15
--model_name	Sentence embedding model	'all-MiniLM-L12-v2'
--label_model	LLM for labeling dialogue state clusters	'open-mixtral-8x22b'
--tau	Minimum transition probability threshold	0.1
--top_k	Number of outgoing edges to retain per node	1
--alpha	Balance between semantic similarity and topology	1.0

---

Citation

If you use this codebase for your research, please cite:

@inproceedings{achref2025conversationalgraph,
  title={A Computational Approach to Modeling Conversational Systems: Analyzing Large-Scale Quasi-Patterned Dialogue Flows},
  author={Mohamed Achref Ben Ammar and Mohamed Taha Bennani},
  conference={IEEE EUROCON 2025 - The 21st International Conference on Smart Technologies},
  year={2025},
  publisher={IEEE},
}

---

Contact

For questions, collaborations, or feedback, feel free to reach out:

• Mohamed Achref Ben Ammar – mohamedachref.benammar@insat.ucar.tn
• Mohamed Taha Bennani – taha.bennani@fst.utm.tn