Naga Language Similarities

Introduction

The Naga are a unique people with distinct sets of cultural and socio values. More than 66 tribes share close cultural ties, a common history, and similar socio-economic practices. They inhabit the mountainous regions of Nagaland, Manipur, Arunachal Pradesh (India) and parts of Sagaing (Myanmar). While each tribe maintains its own identity, the Nagas are united by shared traditions, governance systems, and a deep connection to their land.

The Naga peoples also share something deeper: their languages. Though a person from the Ao tribe often cannot understand someone from the Tangkhul tribe, the words they use for basic concepts like stone, fire, water, and mother reveal striking similarities. These shared roots, called cognates, are traces of common ancestors that computational methods can detect even when the languages have diverged beyond mutual intelligibility.

Languages evolve. New words enter dictionaries every year, old words fade, pronunciations shift. Without a shared written standard, languages diverge. Portuguese and Spanish were once a single tongue; geography split them into separate languages that still echo each other. Naga languages followed the same path over a much longer timescale: mountains and valleys let each community's speech drift its own way.

The data in this study shows that Naga languages trace back to a common origin, branching outward much like the small variations in culture across tribes. The words that survived this divergence, "lung" for stone across 15 languages, "mei" for fire across 4 branches, point to what is shared. That shared core is larger than the differences.

This is the first computational study to measure these connections across 37 Naga languages, using parallel Bible translations as a controlled corpus. The Bible is often the only substantial digital text available in these languages. Because every language translates the same content, differences in character patterns reflect genuine linguistic distance.

The goal is not to rank languages or tribes. It is to find the common threads: the shared words, the similar sounds, the patterns that connect communities across mountains and state borders.

Words That Connect Languages

When you place Naga words for the same concept side by side, the connections become visible. The word for "stone" sounds like "lung" in Ao, "hlung" in Anal, "long" in Konyak, "thlung" in Maring. These are not coincidences. They are the same ancient word, worn into slightly different shapes by centuries of independent change in each community. The table below shows some of the clearest examples, where the shared sounds are immediately recognizable across languages that belong to different branches of the family tree. The full dataset contains many more.

Beyond the words that are spelled similarly, many more Naga words sound alike when spoken even though their spellings differ. For example, "shini" (seven in Tangkhul) and "thenie" (seven in Angami) share the same vowel pattern and rhythm, with only the initial consonant shifted. Similarly, "chishat" (eight in Tangkhul) and "desat" (eight in Zeme) carry the same structure. Written forms can obscure these connections because Naga languages lack standardized orthographies, and different translation teams chose different spelling conventions. When you hear these words spoken, the family resemblance becomes much clearer.

Numbers: The Most Stable Words

Numbers are among the most resistant words to change in any language. They are rarely borrowed and tend to preserve ancient forms. The Naga number systems reveal clear cognate chains across branches:

The word for "five" contains the root "nga/nga" across at least 16 languages and 6 branches. "Three" preserves the root "thum/sum/chum" across Pakan, Zemeic, and Tangkhul-Maring branches. These number cognates are particularly significant because numbers are almost never borrowed between languages; if two languages share a number word, they inherited it from a common ancestor.

🔗 Language Connections

This chord diagram shows which languages share the most vocabulary. Each arc is a language, colored by branch. Ribbons connect languages that frequently use similar words for the same concept. Thicker ribbon = more shared vocabulary. Hover over any arc to highlight that language's connections.

What the Connections Reveal

• Pakan languages (purple arcs) have the thickest internal ribbons. Monsang, Moyon, Chothe, Kharam, Tarao, Anal, Lamkang, and Kom share the densest vocabulary network.
• South Patkaian (blue): Konyak and Phom have the strongest single connection, followed by Chang and Khiamniungan.
• Angami-Pochuri (green): Angami and Chokri are tightest; Sumi and Rengma varieties add breadth to the cluster.
• Cross-branch ribbons show that all branches share some vocabulary with each other. While the within-branch connections are strongest, the ribbons between different branches confirm that Naga languages are not isolated groups but part of one connected family.

The 37 Languages

Branch classifications verified against Glottolog 5.x (Hammarstrom et al., 2024). Five classification errors in initial assignments were corrected using the computational data before verification.

Of these, 24 have complete Bibles (66 books, 1,189 chapters), 10 have New Testament only (27 books, 260 chapters), and 3 are partial. Data coverage ranges from 31,000+ verses (full Bible languages) to just 671 verses (Chuyo, Mark only).

Why Some Languages Are Missing

Several major Naga languages with published Bible translations could not be included. Their translations exist in print but not in any digital format accessible online. We surveyed six platforms (bible.com, Scripture Earth, FCBH, API.Bible, find.bible, Digital Bible Library) and five open NLP datasets (eBible, Taxi1500, BibleMMS, Snow Mountain, Targum) none contained these texts. The Bible Society of India holds copyright on all Naga translations and has not released these specific editions digitally. We invite publishers and community members to help digitize these resources so they can be included in future work.

How We Measured Similarity

Six methods were tested, each capturing a different aspect of what makes languages similar. Three produced strong results; three failed and understanding why they failed is as scientifically valuable as understanding what works.

Text cleaning removing cross-references, verse numbers, English book titles, and footnotes improved Jaccard ARI by 20% and TF-IDF Cohen's d by 27%. One language (Wancho) had 30.6% annotation density inline translator notes explaining Hebrew words. Even after cleaning, its 530,803-word corpus is fully usable.

What the Data Shows

All three primary methods (Jaccard, TF-IDF, Subword JSD) recover known branch-level clusters. The visualizations below use Jaccard, the best method for clustering.

Similarity Heatmap

Each cell is one pairwise similarity score between two languages. Dark = most similar, light = least similar. Languages are sorted by average similarity, so same-branch groups appear as dark blocks along the diagonal.

Jaccard similarity heatmap. Score range: 0.11 to 0.39. The Pakan block is the densest; these 8 languages share the most character trigrams.

PCA: Languages Mapped in 2D

Each dot is a language. Languages that are similar to each other appear close together. The same-colored clusters you see are languages from the same branch, confirming that the computational similarity matches known linguistic classifications. Note that the distances in this chart are compressed: languages on opposite ends of the plot are not as different as they may appear. All Naga languages are closer to each other than any of them are to English, which is excluded from this view. The chart is best read by looking at which languages cluster together, not by measuring distance between clusters.

PC1 and PC2 capture about 30% of total variance. Same-colored dots clustering together = branch classification confirmed by the data.

t-SNE: Nearest Neighbors

This is a different way to visualize the same data. t-SNE focuses on preserving which languages are each other's closest neighbors. If Monsang and Moyon are touching, they are genuinely the most similar pair. Like PCA, the apparent distances between distant groups can be misleading: languages on opposite sides of the chart still share far more with each other than with any language outside the Naga family. The chart shows grouping, not absolute distance.

Axes have no inherent meaning. Only relative proximity matters. Same clusters as PCA, with tighter local grouping.

Extracting Shared Words

Beyond overall similarity scores, we extract specific word translations across all 36 languages using two independent computational methods, then compare them to find high-confidence translations.

Method 1: NPMI (Normalized Pointwise Mutual Information)

For each English word, NPMI measures how strongly it co-occurs with each Naga word in parallel Bible verses. If "water" in English and "tui" in Chothe appear together in verses more often than chance predicts, they're likely translations of each other. NPMI normalizes the score to [-1, 1], solving a problem with standard PMI where rare words that coincidentally co-occur once could score higher than genuine translations.

Threshold: NPMI ≥ 0.25 · Co-occurrence minimum: ≥ 2 verses · Result: ~2,000 English concepts mapped to 36 languages.

Method 2: RAT (Retrieval-Augmented TF-IDF)

A different approach: for each English word, BM25 first retrieves the most relevant Bible verses (like a search engine), then TF-IDF scores Naga words within that focused context. This catches rare, context-specific words that NPMI misses because they appear in too few verses for statistical significance. RAT is fully independent from NPMI no cross-referencing so when both methods agree, the agreement is meaningful.

Where Both Methods Agree

🔍 Word Explorer

Search for any word to see how Naga languages express it. Each colored group = languages that use a similar-sounding word. Same color within a row = same cognate cluster.

Showing sample cognate data. Full dataset (2,000+ words × 36 languages) available in the source repository.

Conclusions & Future Work

What We Found

1. Naga languages are connected character-level patterns reveal shared vocabulary matching known linguistic classifications verified against Glottolog 5.x
2. Jaccard + TF-IDF are the reliable methods for same-script closely-related language comparison; sentence embeddings and unigram frequency fail
3. Text cleaning is essential removing annotation artifacts improved results by 20-27%; embeddings got worse (annotations were their only signal)
4. Two independent word extraction methods (NPMI + RAT) produce 9,087 high-confidence translations where they agree
5. Pan-Naga cognates exist words like lung/stone and mei/fire span 4-6 branches, suggesting deep shared ancestry
6. Five classification errors were corrected by the computational data before verification against linguistic databases

Limitations

• Bible text only everyday vocabulary, songs, stories, proverbs may show different patterns
• Surface orthography, not deep structure cannot capture syntax, morphology, tone, or semantics
• Translation style varies some translations are literal (word-for-word), others dynamic (thought-for-thought)
• Some translations are computationally incorrect not validated by native speakers
• Several major languages are missing their translations are not available digitally
• Orthographic conventions are not standardized spelling variation between translation teams may mask or create artificial similarity

References

• Burling, R. (2003). The Tibeto-Burman languages of NE India. In The Sino-Tibetan Languages. Routledge.
• Cavnar, W.B. & Trenkle, J.M. (1994). N-gram based text categorization. SDAIR-94.
• Cilibrasi, R. & Vitanyi, P. (2005). Clustering by compression. IEEE Trans. Information Theory.
• Hammarstrom, H., et al. (2024). Glottolog 5.0. Max Planck Institute.
• Lu, Y., et al. (2023). Chain-of-Dictionary Prompting. EMNLP 2023. arXiv:2305.06575.
• Marrison, G.E. (1967). The Classification of the Naga Languages. PhD thesis, University of London.
• Morey, S. (2019). The Tangsa-Nocte languages. Himalayan Linguistics, 18(1).
• Singh, A.K. (2006). A study of distance measures for language and encoding identification.
• van Driem, G. (2001). Languages of the Himalayas. Brill.
• Wang, Y., et al. (2023). mPLM-Sim. EMNLP 2023. arXiv:2305.13684.
• Wälchli, B. & Cysouw, M. (2012). Lexical typology through similarity semantics. Linguistics, 50(3).