1 Background and My Purpose
In discussions regarding the relationship between Kurdish and Zazaki, different sides often cite highly speculative rates of similarity or divergence. While some claim that Zazaki is 90% identical to Kurdish, others argue that there is no similarity between the two languages whatsoever. Moving beyond these exaggerated claims that fail to reflect linguistic reality, I developed an online tool capable of measuring the convergence and divergence between the two languages with near-millimetric precision.
Through the Kurdish-Zazaki Comparison Tool I developed, anyone can now examine the relationship between the two languages through clear examples and scientific methods. Rather than relying on ideological assumptions or preconceptions, the system I built is grounded in concrete linguistic data—presenting both their shared heritage and their distinct structural boundaries within a transparent mathematical framework free of artificial manipulation.
2 How Does the Kurdish-Zazaki Comparison Tool Work?
The comparison engine I designed operates on a two-layered architecture based on the principles of computational lexicostatistics and mutual intelligibility:
A. Lexical Module (Lexical Engine) — Step-by-Step Analysis:
To avoid the illusion of superficial letter resemblance when comparing words across both languages, I established a 4-stage filter:
- Phonetic and Orthographic Normalization: First, I equalize equivalent phonemes across both alphabets (
ı ↔ i, ê ↔ e, û ↔ u) and reduce aspirated consonants in Zazaki (ph, th, kh) to their core consonant roots (p, t, k).
- Morphological Suffix and Izafe Stripping: To isolate the lexical root, I strip trailing izafe, feminine, and diminutive endings (
-ê, -a, -e, -ya, -ye, -ıkhe) to reach the bare stem.
- Historical Iranic Sound Shift Mapping: I test for regular phonological correspondences across Northwestern Iranic languages (
j/c/d/g ↔ z, b/w/m ↔ v, o/u ↔ a) as well as initial h- prosthesis (e.g., ewr ↔ hewr).
- Consonant Skeleton and Protected Levenshtein Calculation: By stripping vowels, I measure the consonant skeleton and subsequence ratio. To prevent accidental single-letter overlaps in short words from generating unfair scores, I enforce a strict initial-sound (onset) shield.
B. Sentence Module (Syntactic & Semantic Engine) — Content-Backbone and Cumulative Shift Model:
Assigning equal weight to every word in a sentence comparison is scientifically misleading; two sentences do not become mutually intelligible merely because pronouns or prepositions like
ez, tu, min, xo/xwe match. Therefore, I developed the
"Content-Backbone" model:
- Backbone Weighting: While auxiliary elements (pronouns, prepositions, conjunctions) receive a weight of 0.25, core meaning-bearing nouns, adjectives, interrogatives, and verbs receive 1.00 (Full Backbone Weight). If no core content words match between two sentences, the sentence score is strictly 0% even if the pronouns are identical.
- Copula Balancing: Because predicative copulas (
ye / e) are written as separate tokens in Kurdish but attached as suffixes (-o / -yo / -ya) in Zazaki, I normalize them to prevent artificial inflation of the word-count denominator.
- Cumulative Sound Shift Rule: A sound shift in a single word does not break mutual intelligibility (e.g., in
Daê non bıde mı ↔ Daê nan bide min, the exact matches Daê and bıde/bide provide an auditory anchor for the listener, making the sentence 100% intelligible despite the non/nan vowel shift). However, when every single word in a sentence undergoes simultaneous phonetic or morphological modification—as in Navê te çi ye ↔ Namê to çıko (nav→nam, te→to, çi ye→çıko)—leaving zero 100% identical anchor words, these differences accumulate and create auditory opacity. The engine detects this cumulative shift and lowers the sentence from 88% to the 38% (Moderate) tier.
3 How Did I Select the Lexical and Sentential Corpus?
I did not gather the compared materials at random; to ensure linguistic impartiality, I bound the corpus to the following objective filters:
-
Swadesh Core Vocabulary Lists: The universal Swadesh lists—used in comparative and historical linguistics to measure genetic relatedness—form the core backbone. This list comprises concepts impervious to ideological manipulation, such as natural phenomena, body parts, basic actions, numerals, and directions.
-
Natural Terms and Everyday Domains: To reflect organic contact at the grassroots level, I included traditional vocabulary that has lived for millennia: animals, fruits, vegetables, agricultural and craft tools, and household items.
-
Exclusion of Artificial and Recent Loanwords: Shared foreign loanwords (from Turkish, Arabic, or Western languages) that entered both languages in modern political, philosophical, or technological contexts are excluded from the core dataset so as not to distort the authentic profile of either language; native equivalents are prioritized.
-
Natural Sentence Syntax: Rather than artificial word-for-word translations forced from one language into the other, I selected natural everyday sentences with authentic verbal conjugation and syntax in both languages.
4 How Are Similarity and Divergence Rates Calculated? (Where Do 88%, 38%, and 18% Come From?)
Rather than assigning arbitrary percentages, I established a 5-Tier Scoring System grounded in concrete linguistic criteria. Here is the mathematical explanation behind the question users ask most frequently: "Where do the numbers 88%, 38%, and 18% come from?"
📌 Why 88%, 38%, and 18%? (The Band Centroid Principle):
In linguistic comparison, similarity between two items spans a continuous spectrum from 0 to 100. However, raw Levenshtein percentages (such as 66.6%, 71.4%, or 42.8%) produce artificial fluctuations depending on word length. To eliminate this distortion, I applied the statistical "Band Centroid (Representative Midpoint)" method:
• Midpoint of the 75% – 99% band (High intelligibility with a single regular sound shift) → 88%
• Midpoint of the 26% – 50% band (Morphologically diverged historical cognates) → 38%
• Midpoint of the 10% – 25% band (Weak overlap sharing only an initial consonant or distant resonance) → 18%
Consequently, when a 3-letter word and an 8-letter word exhibit the exact same linguistic phenomenon, they do not receive different scores due to character count; they converge in the same scientific tier.
100% Exact Equivalence (Identical Shared Root)
Tier 1
Words and sentences whose root consonants and vowel sequences match identically after phonetic normalization, exhibiting no sound shifts or affixation divergence.
How It Is Calculated: When the Levenshtein distance of the consonant and vowel skeleton is 0 after orthographic equalization (ı=i, ê=e, ph=p, kh=k), a 100% score is registered. For sentences, 100% is awarded when the consonant skeleton of core backbone words matches completely and the weighted average is ≥85.
Examples: dest ↔ dest | gıran ↔ giran | Daê non bıde mı ↔ Daê nan bide min
88% Close Cognate Rate (Regular Sound Shift)
Tier 2
Sister words that follow historical Northwestern Iranic sound laws (e.g., j ↔ z, b ↔ v, m ↔ v, o ↔ a, u ↔ a) or initial h- prosthesis, and are immediately intelligible to everyday speakers with only a single sound or izafe difference.
How It Is Calculated: When the consonant skeleton is preserved and the difference rests on a single regular phonetic shift, a feminine -e ending, or a Levenshtein similarity of ≥70%, the score is assigned as 88% (the centroid of the 75%–99% band).
Examples: dare ↔ dar | phonc ↔ pênc | roce ↔ roj | name ↔ nav
38% Diverged Cognate Rate (Distant Sister Roots)
Tier 3
Roots that share an undeniable etymological origin but have undergone morphological or phonetic divergence over time—making them difficult to recognize instantly in casual speech—as well as sentences affected by cumulative sound shifts across all tokens.
How It Is Calculated: Pairs with a consistent initial consonant or core consonant subsequence but two or more internal sound divergences, as well as sentences where every word undergoes simultaneous phonetic/affix modification, are evaluated at 38% (the centroid of the 26%–50% band).
Examples: dımpıstıkhe ↔ dûpişk | goş ↔ guh | Namê to çıko ↔ Navê te çi ye
18% Weak Similarity Rate (Coda Divergence)
Tier 4
Weak pairs that share only a distant acoustic resonance, a single root consonant, or just one shared word within a longer sentence, producing no mutual intelligibility.
How It Is Calculated: When the final consonant (coda) diverges in short words, phonetic similarity remains in the 30%–49% boundary, or only 1 content word matches in a sentence of 3+ content words, the score is capped at 18% (the centroid of the 10%–25% band).
Examples: khoçhıkhe ↔ kevçî | şıt ↔ şîr
0% Complete Divergence Rate (Unrelated / Distinct Root)
Tier 5
Words whose roots, consonant skeletons, and etymological origins are completely unrelated, or sentences that share no core content words.
How It Is Calculated: My algorithmic firewall intervenes; regardless of shared infinitive endings, pronouns, or accidental letter overlaps, once root divergence is detected, the score is strictly set to 0%.
Examples: çê / keye ↔ mal | sole ↔ xwê | şıma ↔ hûn
5 How Did I Select Dialectal Variants in Both Languages?
Both Kurdish (Kurmanji) and Zazaki (Kirmancki/Dimili) possess rich internal dialectal diversity. To establish a fair baseline for comparison, I followed a two-step method when selecting variants:
Rule 1 (Prevalence in the Source Language):
When determining the variant of a word, I primarily selected the form most widely accepted and standardized in that language's written literature and among its speakers. Marginal forms restricted to a narrow locality were not chosen as general representatives.
Rule 2 (Proximity Balance):
Where multiple legitimate and widespread variants satisfy the first rule within a language, I selected the natural variant that stands phonetically closer in order to accurately reflect historical relatedness. This choice is not an artificial approximation, but an objective effort to make visible the historical bond preserved within the natural richness of both languages.
6 Phonemic Reality Over Alphabetical Conventions
Writing systems are conventions created later by humans, whereas the essence of a living language lies in its sounds. Therefore, I eliminated artificial divisions arising merely from alphabetical or spelling differences:
- Orthographic Equalization: When the letter i in the Kurdish alphabet and the letter ı in the Zazaki alphabet represent the exact same phoneme, this is an alphabetical preference rather than a linguistic difference. My system treats these sounds as equal (e.g., Kurdish
bira and Zazaki bıra are phonemically identical).
- Aspirated Consonants: Aspirated digraphs stemming from the phonological richness of Zazaki (ph, th, kh) are analyzed by reducing them to their core consonant roots (p, t, k) when compared with their Kurdish counterparts.
- Preservation of Morphological Identity: Living feminine endings or oblique/ergative inflections in Zazaki are not artificially truncated; the authentic identity of each word is preserved.
7 Scientific Safeguards, Data Discipline, and Expert Override
The only way to remain impartial when comparing Kurdish and Zazaki is to bind every evaluation to algorithmic safeguards. While designing this system, I strictly enforced the following principles:
Just as I did not cherry-pick shared words to artificially inflate the similarity rate, I did not intentionally foreground divergent words to portray the languages as completely disconnected. Using the internationally standardized Swadesh list as the core backbone rules out subjective manipulation from the outset.
🔒 Human-in-the-Loop Calibration (Expert Override):
In computational linguistics, no purely mathematical string algorithm can flawlessly capture suppletion (where verb roots change completely across tenses, e.g., vatene ↔ gotin), historical metathesis (sound transposition, e.g., vêr ↔ ber), or idiomatic shifts 100% of the time. In exceptional cases where character sequences mislead the automated calculation, I implemented an Expert Calibration (Expert Override) mechanism in accordance with standard linguistic ethics. This mechanism never invents arbitrary numbers (such as 65% or 92%); it simply locks the linguistic reality overlooked by the string algorithm into its rightful place among the 5 constitutional tiers (100%, 88%, 38%, 18%, 0%) defined above.
8 Conclusion and My Call for Contributions
The Kurdish-Zazaki Comparison Tool I developed is not a static dictionary; it is a dynamic, living corpus that continuously grows through the contributions of speakers and researchers.
To further enrich and refine this corpus, I invite everyone to submit word and sentence pairs. Ensuring that submitted entries comply with the methodological rules outlined above (especially single-word entry in lexical mode, accurate orthography, and semantic equivalence) is vital for the academic integrity of this work.
All words and sentences submitted by visitors are forwarded directly to my administration panel, where they are verified through the scoring engine and my editorial review before publication.
I developed this tool with a completely transparent approach open to public scrutiny. As is often the case in language technologies of this kind, algorithmic calculations or word/sentence alignments may occasionally produce inaccurate values or deviations that violate the methodological principles I detailed above. If you come across such errors, I kindly ask you to alert me. More broadly, I would be delighted to hear any suggestions, contributions, or objections you may have. You can share your thoughts in the comments section below or contact me directly.
Cemal Atila
October 15, 2026
My Contact Information:
Email: cemal.zkm@gmail.com
My Facebook account
My Instagram account
My Twitter (X) account
📋 Version and Update Log (Changelog)
v1.0 — October 15, 2026
Launched the Kurdish-Zazaki-Persian Lexical (446+ core concepts) and Sentential (93+ natural sentences) Comparison Engine, 5-Tier Lexicostatistical Scoring System, Content-Backbone Model, Cumulative Sound Shift Rule, and Expert Override infrastructure.
💬 Feedback, Suggestions, and Objections (0)
✍️ Leave a Comment, Suggestion, or Objection (No email or login required)
No approved comments yet. Be the first to share your thoughts!