Cross-Linguistic Institutional Grammar Analysis

Testing whether the ADICO institutional grammar framework is linguistically universal or typologically specific.

Overview

This repository contains code, data, and analysis for a research project examining how linguistic typology shapes institutional reasoning. We test the Crawford-Ostrom Institutional Grammar (IG) framework by generating deliberative debates in four typologically diverse languages and analyzing whether the ADICO structure (Attribute-Deontic-Aim-Condition-Or-else) emerges naturally.

Research Question

Does the "attributed, deontically qualified action" assumed by institutional analysis represent a universal cognitive unit, or does it reflect the grammatical affordances of English and similar languages?

Key Finding

ADICO is a genre-specific grammar, not a universal template. All four languages can produce ADICO-compatible outputs when explicitly instructed, but none do so by default. Each language channels normativity through different grammatical pathways:

Language	Alignment	Default Normative Strategy
English	Nominative-accusative	Agent-action framing, modal deontics
Basque	Ergative-absolutive	Process-orientation, distributed agency
Czech	Nom-acc + aspect	Middle voice, state descriptions
Hebrew	Nom-acc + binyanim	Causative templates, implicit deontics

Repository Structure

ErgativeAgentsSims2025/
├── debate.py                 # Main debate generation script
├── research_agent.py         # Cross-linguistic analysis engine
├── visualizations.py         # Chart and dashboard generation
├── requirements.txt          # Python dependencies
│
├── logs2025/                 # Raw debate transcripts (JSONL)
│   ├── english_*.jsonl
│   ├── basque_*.jsonl
│   ├── czech_*.jsonl
│   └── hebrew_*.jsonl
│
├── research_outputs/         # Analysis results
│   └── session_*/
│       ├── reports/          # JSON + Markdown reports
│       └── visualizations/   # Charts and dashboards
│
├── article_figures/          # Publication-ready figures
│   ├── appendix/             # Appendix visualizations
│   └── *.png
│
├── docs/                     # Documentation
│   ├── ARTICLE_MATERIALS.md  # Draft article content
│   ├── THREE_STUDY_COMPARISON.md
│   ├── DATA_DICTIONARY.md    # Data file documentation
│   └── METHODOLOGY.md        # Full methodology
│
├── analyzers/                # NLP analysis modules
│   ├── morphological_analyzer.py
│   ├── syntactic_analyzer.py
│   └── ...
│
└── tests/                    # Test suite

Installation

Prerequisites

• Python 3.10 or higher
• OpenAI API key (for debate generation)
• ~4GB disk space (for NLP models)

Setup

# Clone the repository
git clone https://github.com/yourusername/ErgativeAgentsSims2025.git
cd ErgativeAgentsSims2025

# Create virtual environment
python -m venv venv

# Activate (Windows PowerShell)
.\venv\Scripts\Activate.ps1

# Activate (Linux/Mac)
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Download NLP models
python -m spacy download en_core_web_sm
python -c "import stanza; stanza.download('eu')"  # Basque
python -c "import stanza; stanza.download('cs')"  # Czech  
python -c "import stanza; stanza.download('he')"  # Hebrew

# Configure API key
echo "OPENAI_API_KEY=your_key_here" > .env

Quick Start

Reproducing the Analysis

To reproduce the main findings using existing data:

# Analyze the neutral condition debates (4 languages)
python research_agent.py --logs logs2025/english_ai_harm_prevention_open_20260210_*.jsonl \
                              logs2025/basque_ai_harm_prevention_open_20260210_*.jsonl \
                              logs2025/czech_ai_harm_prevention_open_20260210_*.jsonl \
                              logs2025/hebrew_ai_harm_prevention_open_20260210_*.jsonl

# Results appear in research_outputs/session_YYYYMMDD_HHMMSS/

Generating New Debates

# Run a 6-round debate with neutral prompts
python debate.py --language english --open-form --topic ai_harm_prevention --rounds 6

# Run with rule-demanding prompts (15 rounds)
python debate.py --language basque --open-form --topic ai_harm_prevention --rounds 15 \
                 --prompt-style proposal

# Run with anti-rules prompts
python debate.py --language czech --open-form --topic ai_harm_prevention --rounds 6 \
                 --prompt-style anti-rules

Generating Figures

# Generate appendix figures
python generate_appendix_figures.py

# Generate article figures  
python generate_article_figures.py

# Figures saved to article_figures/

Data Description

Debate Logs (logs2025/)

JSONL files containing debate transcripts. Each line is a JSON object with:

{
  "round": 1,
  "speaker": "Agent_A",
  "timestamp": "2026-02-10T13:35:56.123Z",
  "content": "Debate utterance text...",
  "metadata": {
    "language": "english",
    "topic": "ai_harm_prevention",
    "prompt_condition": "neutral"
  }
}

File naming convention: {language}_{topic}_{mode}_{datetime}_{hash}.jsonl

Research Reports (research_outputs/)

Analysis outputs include:

• research_report_*.json - Structured quantitative data
• research_report_*.md - Human-readable analysis
• SESSION_SUMMARY.md - Cross-linguistic comparison

Key metrics in JSON reports:

• subjects_per_sentence - Agent visibility measure
• hhi_agency - Herfindahl-Hirschman Index for agency concentration
• voice_valency_analysis - Distribution of grammatical voice types
• information_status - Du Bois Given A Constraint metrics

See docs/DATA_DICTIONARY.md for complete documentation.

Methodology

Three-Condition Experimental Design

Condition	Prompt Style	Rounds	Purpose
Rule-demanding	"formulate clear rules and guidelines"	15	Test ADICO capacity
Anti-rules	"focus on experiences, not formal rules"	6	Test natural defaults
Neutral	"describe how things should be"	6	Baseline comparison

Analytical Framework

1. Subject Realization - Measures explicit agent naming (subjects per sentence)
2. Agency Distribution (HHI) - Concentration of grammatical roles (A/S/O)
3. Voice Analysis - Distribution of active, passive, middle, causative constructions
4. Information Status - Du Bois's Given A Constraint adherence
5. ADICO Failure Modes - Which components resist coding, by language

Methodological Constraints

Results should be interpreted with awareness that:

• Debates are LLM-mediated (GPT-4o), not native speaker production
• Single topic (AI harm prevention) may not generalize
• NLP tools have variable accuracy across languages
• Genre conventions may influence outputs

See docs/METHODOLOGY.md for full methodology documentation.

Key Results

Quantitative Summary (Neutral Condition)

Metric	English	Basque	Czech	Hebrew
Subjects/Sentence	1.78	0.80	1.80	1.37
HHI Agency	0.415	0.457	0.336	0.336
Active Transitive %	81.5	23.5	53.5	21.3
Given A Adherence %	51.9	42.8	54.1	50.1

Emergent Grammars

The analysis identifies four deliberative grammar types that emerge naturally:

1. PPO (Process-Participant-Orientation) - Basque default
2. RST (Relational-State-Transition) - Czech default
3. AFIG (Affected-First Grammar) - Ergative-aligned discourse
4. ECG (Enunciative-Contextual Grammar) - Hebrew default

Citation

If you use this project in research, please cite:

@software{cross_ling_ig_2026,
  title = {Cross-Linguistic Institutional Grammar Analysis},
  author = {[Author Name]},
  year = {2026},
  url = {https://github.com/yourusername/ErgativeAgentsSims2025},
  note = {Code and data for testing ADICO universality across languages}
}

License

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

Acknowledgments

• OpenAI GPT-4o for debate generation
• Stanford NLP Group for Stanza
• spaCy for English analysis
• The Ostrom Workshop for institutional analysis frameworks

Related Work

• Crawford, S., & Ostrom, E. (1995). A Grammar of Institutions. APSR
• Dixon, R. M. W. (1994). Ergativity. Cambridge University Press
• Du Bois, J. W. (1987). The Discourse Basis of Ergativity. Language
• Dowty, D. (1991). Thematic Proto-Roles. Language

Contact

For questions about the code or methodology, please open an issue on GitHub.