CEIEF — Cultural Education Interaction Enhancement Framework

文化教育交互增强框架

An AI-assisted interaction framework for cultural education, integrating a Cultural Knowledge Graph (KG), Large Language Model (LLM), and an Agent-based Interaction Controller to achieve cultural context retention, semantically coherent generation, and adaptive interaction regulation.

---

Overview

CEIEF addresses three research questions:

• RQ1: Can a knowledge graph improve semantic coherence and cultural grounding in cultural learning dialogues?
• RQ2: Can an agent control mechanism enhance adaptivity and pedagogical regulation in multi-turn interaction?
• RQ3: Does the full CEIEF framework outperform ablated configurations across cultural semantic coherence, cognitive depth activation, and interaction adaptivity?

Architecture

┌──────────────────────────────────────────────┐
│            CEIEF Three-Layer Architecture     │
├──────────────────────────────────────────────┤
│  Layer 3: Agent Controller (Adaptive Layer)  │
│    - State tracking: s_t = [g_t;u_t;m_t;b_t]│
│    - Action space: 7 pedagogical actions     │
│    - Rule-based policy with diversity         │
├──────────────────────────────────────────────┤
│  Layer 2: LLM Dialogue Generator             │
│    - Prompt package assembly                 │
│    - KG evidence injection                   │
│    - Role-consistent generation              │
├──────────────────────────────────────────────┤
│  Layer 1: Cultural Knowledge Graph           │
│    - 300 cultural triples                    │
│    - 5 entity types, 12 relation types       │
│    - Neo4j + CSV dual backend                │
└──────────────────────────────────────────────┘

Experimental Conditions

Condition	Configuration	Purpose
llm_only	LLM only	Baseline
kg_llm	KG + LLM	Test KG contribution
llm_agent	LLM + Agent	Test Agent contribution
full_ceief	KG + LLM + Agent	Complete framework

---

Installation

git clone https://github.com/chengYu23/CEIEF.git
cd CEIEF
pip install -r requirements.txt

Configuration

# Copy example environment file
cp .env.example .env

# Edit .env with your actual credentials
# ANTHROPIC_API_KEY=your_api_key
# NEO4J_URI=bolt://localhost:7687
# NEO4J_PASSWORD=your_password

The config.yaml reads credentials from environment variables. Do not hardcode secrets.

---

Usage

1. Load Knowledge Graph to Neo4j (optional)

python scripts/load_kg_to_neo4j.py

If Neo4j is unavailable, the system automatically falls back to CSV-based keyword retrieval.

2. Run Experiments

# Full CEIEF (KG + LLM + Agent)
python scripts/run_full_ceief.py --task_id T001 --turns 5

# KG + LLM
python scripts/run_kg_llm.py --task_id T001 --turns 5

# LLM + Agent
python scripts/run_llm_agent.py --task_id T001 --turns 5

# LLM-only baseline
python scripts/run_llm_only.py --task_id T001 --turns 5

Task IDs: T001–T020 (see data/tasks/task_bank.json)

3. Evaluate Results

# Extract automatic metrics from session logs
python scripts/extract_metrics.py --log_dir logs/raw_dialogues/ --verbose

# Generate per-condition summary statistics
python scripts/summarize_results.py

---

Knowledge Graph

300 simulated cultural knowledge triples covering:

• Traditional festivals: 端午节, 春节, 中秋节, 清明节, 重阳节, 七夕, 元宵节
• Cultural values: 孝道, 礼仪, 集体记忆, 文化认同, 和谐, 中庸之道
• Historical figures: 屈原, 孔子, 孟子, 花木兰, 王昭君
• Philosophical traditions: 儒家思想, 道家思想, 佛教
• Lifecycle practices: 成人礼, 祭祖, 婚礼

Entity types: cultural_concept, historical_actor, event_practice, spatiotemporal_context, value_orientation

Relation types: embodies, commemorates, belongs_to, contrasts_with, transfers_to, interprets, originated_in, practiced_by, associated_with, reinforces, challenges, occurs_in

---

Evaluation Metrics

Automatic

Metric	Symbol	Description
Cultural Semantic Coherence	CSC	KW density + topic overlap + KG citation
KG Coverage Ratio	KGCR	Proportion of retrieved entities cited in response
Response Depth Score	RDS	Length + depth vocabulary + sentence count
Action Diversity Index	ADI	Shannon entropy of action history

Human (5-point Likert, 3 raters)

• Historical Context Matching
• Cultural Label Alignment
• Language Style Modal Alignment

Target inter-rater reliability: Cohen's κ ≥ 0.70

License

This project is licensed under the MIT License — see the LICENSE file for details.