Elamite Dataset Lab

Word2Vec Network Analysis of Ancient Elamite Texts

UntN-Nasu Collection | Linguistic Pattern Discovery | Phrase Boundary Detection

---

Overview

This project applies Word2Vec word embeddings and network graph analysis to a corpus of Elamite texts to explore how words form constructions and identify patterns that may assist with lemmatization and linked data representation.

Corpus Statistics
85 Documents	2,582 Tokens
649 Unique Words	1,397 Combined Edges

---

Bi-gram + Similarity Network Analysis

Methodology

We combine two types of linguistic relationships:

Relationship Type	What it captures	Edge Weight
Syntagmatic (Bi-gram)	Words appearing adjacent in text	Low (0.15 per occurrence)
Paradigmatic (Similarity)	Words with similar distributions	High (cosine × 1.5)
Reinforced (Both)	Fixed constructions	Combined

Combined Network

Cyan = bi-gram edges, Red = similarity edges, Gold = reinforced (both)

---

Fixed Constructions (Reinforced Edges)

32 word pairs that are BOTH adjacent AND distributionally similar

---

Centrality Analysis

Eigenvector = connected to important words | Bridging = connects different clusters

---

Edge Type Analysis

Distribution of syntagmatic vs paradigmatic relationships

---

Document Stylometry (PCA)

Documents positioned by vocabulary usage patterns

---

Key Findings

1. Fixed Constructions Identified

Word Pair	Similarity	Interpretation
(d)in-šu-uš-na-ak — a-ak	0.89	Divine name + connective
dingir-gal — u2-me	0.81	"Great god" + verbal marker
ku-ši-ih — (d)in-šu-uš-na-ak	0.81	Grammatical + divine name
si-it-me — u2-me	0.77	Verbal forms clustering

2. Clause Boundary Candidates

Word	Bridging Score	Hypothesis
ku-ši-ih	93.9	Connects different clause types
a-ak	83.0	Conjunction — joins clauses
u2-me	80.4	Verbal element

3. Royal Title Formula

(m)un-taš-dingir-gal  ša-ak  an-za-an  šu-šu-un-ka  su-un-ki-ik
[Royal name]          [?]    [Anshan]  [of Susa]    [king]

This bi-gram chain appears 66+ times — the standard royal titulary.

4. The -me Suffix Class

Words ending in -me form a tight distributional cluster, suggesting a verbal/participial marker:

• u2-me, si-it-me, ta-ak-me, hu-us-si-ip-me

---

Original Network Analysis

Full Word Network

Node size = connectivity, Color = determinative type

---

Core Hub Words

---

Word Similarity Heatmap

---

Morphological Distribution

---

Interactive Visualizations

File	Description
visualize_network.html	Original similarity network explorer
visualize_bigram_network.html	Bi-gram + similarity network with edge type filtering

Open in browser to:

• Toggle between edge types (bi-gram, similarity, reinforced)
• Size nodes by eigenvector or bridging centrality
• Search for specific words
• Click nodes to explore connections

---

Project Structure

ElamiteDatasetLab/
│
├── Data & Models
│   ├── texts/                          # 85 document text files
│   ├── UntN-Nasu texts Word-level.csv  # Source data
│   └── elamite_word2vec.model          # Trained Word2Vec model
│
├── Network Analysis (Similarity)
│   ├── edges_similarity.csv            # Similarity edge list
│   ├── nodes_attributes.csv            # Node attributes
│   └── elamite_graph.json              # Full similarity graph
│
├── Network Analysis (Bi-gram + Similarity)
│   ├── bigram_similarity_edges.csv     # Combined edge list
│   ├── nodes_centrality.csv            # Eigenvector & bridging scores
│   ├── document_pca.csv                # Document stylometry
│   └── bigram_similarity_graph.json    # Combined graph
│
├── Visualizations
│   ├── visualize_network.html          # Interactive (similarity)
│   ├── visualize_bigram_network.html   # Interactive (combined)
│   ├── bigram_similarity_network.png   # Combined network
│   ├── reinforced_constructions.png    # Fixed constructions
│   ├── centrality_comparison.png       # Eigenvector vs bridging
│   ├── edge_type_analysis.png          # Edge distribution
│   ├── document_pca_space.png          # Stylometry
│   ├── network_overview.png            # Full network
│   ├── network_hub_subgraph.png        # Hub words
│   ├── similarity_heatmap.png          # Similarity matrix
│   └── morphological_distribution.png  # Suffix stats
│
├── Scripts
│   ├── generate_txt_files.py           # CSV → text files
│   ├── run_word2vec.py                 # Train Word2Vec
│   ├── analyze_embeddings.py           # Clustering analysis
│   ├── build_network_graph.py          # Similarity network
│   ├── bigram_similarity_network.py    # Combined network
│   ├── create_visualizations.py        # Similarity visualizations
│   └── create_bigram_visualizations.py # Combined visualizations
│
├── Reports
│   ├── LINGUISTIC_CONCLUSIONS.md       # Deep analysis findings
│   ├── bigram_similarity_report.txt    # Combined network report
│   ├── network_analysis_report.txt     # Similarity network report
│   └── Elamite_Word2Vec_Report.md      # Methodology report
│
└── Linked Data
    └── elamite_triples.nt              # RDF N-Triples

---

Usage

# 1. Generate text files from CSV
python3 generate_txt_files.py

# 2. Train Word2Vec model
python3 run_word2vec.py

# 3. Run embedding analysis
python3 analyze_embeddings.py

# 4. Build similarity network
python3 build_network_graph.py

# 5. Build bi-gram + similarity network
python3 bigram_similarity_network.py

# 6. Generate all visualizations
python3 create_visualizations.py
python3 create_bigram_visualizations.py

Requirements:

pip install gensim scikit-learn numpy matplotlib networkx

---

Linked Data Export

Ready for knowledge graph representation:

• bigram_similarity_edges.csv — Import into Gephi with edge type column
• elamite_triples.nt — RDF N-Triples format
• nodes_centrality.csv — Node attributes for SPARQL

---

Team

Parsa Faraji & Adam Anderson

---

License

For academic use.