A Few Words on Zazalingo

Throughout my years of research on the Zaza language, the most fundamental deficiency I observed was our language's defenselessness in the face of digital technologies. This vulnerability threatens our mother tongue from multiple angles. On one hand, attempts to dismiss our language as merely a dialect and dissolve it within neighboring tongues continue unabated. On the other hand, artificial intelligence models trained on unvetted, noisy web data blend Zazaki with surrounding languages, producing a distorted, barely recognizable version of our tongue.

Inherited from our ancient ancestors, endowed with thousands of years of oral heritage, rich expressive diversity, and a distinct grammatical architecture, our language is unfortunately losing ground day by day in a rapidly moving digital world. UNESCO reports already classify Zazaki as an endangered language; compounding this vulnerability, rapid deformations introduced by generative AI make the picture even grimmer. For the past two years, I searched restlessly for ways to reverse this trend. Following numerous dead ends, my vision crystallized in late 2025, and I immediately set to work. After an intense and rigorous year of labor, here in late 2026, I present Zazalingo to you. How does Zazalingo work, what purpose does it serve, and how will it benefit our language? Allow me to answer these questions in detail.

How Do Modern Translation Systems Work?

Before exploring the operational principles of Zazalingo, it is necessary to briefly understand how contemporary AI translation systems function. Modern neural machine translation (NMT) architectures and large language models (LLMs) statistically and semantically analyze vast corpora composed of millions of sentence pairs. The system does not translate word-by-word like a conventional dictionary; it evaluates entire sentences within context. It models probabilistic relationships to determine which word appears in which syntactic setting, tense, and relation. When insufficient direct data exists between two languages, systems bridge the semantic gap through global pivot languages, typically English. However, these massive architectures suffer from a critical flaw: if the underlying training corpus lacks robust, verified, and curated data for a given language, the system inevitably hallucinates or forcibly projects the grammar and vocabulary of foreign languages onto it. This is precisely what happens to Zazaki, which is classified in computational linguistics as a low-resource language.

How Was Zazalingo Engineered and How Does It Operate?

To overcome this impasse, initiative had to be reclaimed from uncontrolled web models. In Zazalingo, large language models are not deployed as autonomous translators; rather, they serve strictly as computational inference engines executing rules and verified vocabulary curated by us.

In developing Zazalingo, I addressed the low-resource language challenge through a dual-layer hybrid architecture. First, I manually compiled a core database consisting of thousands of validated parallel sentences and words across Zazaki, Turkish, and English. Designated as the "Golden Corpus" (Altın Bellek), this dataset forms the immutable reference backbone of the system. In the subsequent phase, I applied synthetic expansion to the sentences in the Golden Corpus. Synthetic expansion is a technique in low-resource computational linguistics where grammatical frames are systematically expanded by alternating subject pronouns, negation, interrogative moods, and temporal adverbs while preserving structural integrity.

Having expanded the Golden Corpus through synthetic augmentation, I established the second pillar of the platform. No matter how large an empirical dataset grows, anticipating every exact phrasing a human might generate is impossible; static databases alone remain inherently incomplete. At this stage, rather than leaving generative AI unconstrained, we activate Example-Based Machine Translation (EBMT) and Few-Shot Retrieval-Augmented Generation (RAG). When a user requests a translation not directly present in the Golden Corpus (such as "Today I am going to the city"), the system detects the closest validated syntactic frame: "Today I am going to the village" (Ez ewro şona dewe). It then retrieves the word for "city" (suke) from the verified lexicon and instructs the model: "Preserve the reference frame's verb inflection and tense aspect exactly; replace 'village' with 'city'." Reasoning by analogy over this certified frame, the model produces "Ez ewro şona suke." The model invents nothing spontaneously; it derives its output strictly from verified structures and vocabulary.

What happens when the query has no direct analog in the Golden Corpus? Suppose a user enters "Do you have children?", a structure not yet recorded in the reference memory. In such cases, the system initiates morphological decomposition and lemmatization, stripping inflectional affixes down to raw lexical roots:

These roots are matched against validated entries in the Golden Lexicon (child → doman). Finally, these morphemes are assembled via pre-programmed Zazaki genitive, plural, and interrogative templates:

Naturally, this pipeline is not immune to anomalies; rare vocabulary or foreign elements may occasionally lead to an inaccurate derivation, and the intricate split-ergative alignment of Zazaki presents ongoing challenges. Nevertheless, as corrected translations undergo editorial review and enter the Golden Corpus, the system systematically closes these gaps and compounds its accuracy over time.

Alphabetical Discipline

A major obstacle in contemporary Zazaki written literature is the haphazard coexistence of competing orthographies. This presents several challenges: First, the Bedirxan alphabet, conceived specifically for the phonology of Kurdish, falls short of accurately capturing the rich phonetic inventory of Zazaki. Persistent efforts to force Zazaki into that script lead to substantial phonetic distortions. Second, reactions against such misalignments have spawned divergent ad-hoc orthographic traditions (which cannot truly be deemed standardized systems). For instance, the closed /e/ is inconsistently rendered as either ê or é. Third, writers frequently mix and match multiple orthographies within single texts.

For Zazalingo's orthographic foundation, I adopted the Jacobson/Ware alphabet, which represents the phonetic reality of Zazaki with precision and academic integrity. The underlying algorithms strictly enforce this standard: every generated output adheres uniformly to the standard 32-letter + 4 digraph Latin alphabet for Zazaki.

Preservation of Dialects and Variant Diversity

While standardization is invaluable in orthography, it becomes counterproductive when applied dogmatically to regional linguistic diversity. For this and numerous related reasons, I firmly oppose forced "Standardization of Zazaki." Monolithic standardization constitutes an ideological homogenization that threatens the organic ecosystem evolved over millennia. When designing Zazalingo, I categorically rejected reductionist paradigms that elevate one regional dialect as the sole "standard" while delegitimizing all others as "erroneous."

The more equitable and sustainable model is this: maintain strict collective orthographic discipline while cultivating a pluricentric language model where all regional expressions flourish. While one native speaker says awe for water, another naturally uses awke, owke, or ağwe; the phrase "Tı zaf rında" is no more or less legitimate than "Tı zaf hola." To the extent of my knowledge across regions, I have incorporated these variant expressions directly into the engine.

It must be acknowledged that Zazalingo initially relies primarily on Northern Zazaki, with comparatively modest coverage of Central and Southern dialects. As Northern Zazaki is my native variant, the initial core was built upon it. Verified elements from Central and Southern Zazaki have been integrated, yet their volume remains limited. Through community collaboration, we will transcend this limitation: as speakers of Central and Southern dialects utilize the system and submit suggestions, Zazalingo will progressively reflect the full spectrum of our language.

A Universal Bridge to World Languages

My objective in creating Zazalingo was not merely to establish translation between Zazaki and Turkish, but to connect Zazaki directly with world languages. Today, the Zaza diaspora spans the globe, interacting continuously with diverse societies. Direct translation bridges between major world languages and Zazaki facilitate everyday communication for our diaspora while elevating our language's international visibility.

Presently, Zazalingo supports bidirectional translation between Zazaki and English, German, French, Dutch, Spanish, Italian, Portuguese, Russian, Arabic, Persian, Kurdish, Chinese, Japanese, and Greek, alongside Turkish. Expanding this scope to additional global languages is purely a function of time and computational resources. While I evaluate translation quality directly in languages I speak, broader validation relies on intermediate pivots (primarily English). Insight and feedback from multilingual Zazas will be vital in refining these multilingual paths.

The Vital Role of Community Contribution

Zazalingo was conceived not as a static software package, but as a living organism nourished by the collective contributions of the Zaza community. The framework is ready to accommodate the full depth of our language, but regional diversity must be supplied by our people. Here, speakers of Central and Southern dialects hold a crucial responsibility.

To facilitate this, I positioned a Community Contribution module directly beneath the translation interface. Any speaker can propose regional vocabulary or phrasing. To maintain linguistic purity and prevent noise, submissions do not enter the live system directly; they are routed to the administrative console and merged into the Golden Corpus only following rigorous editorial and orthographic verification.

Technological Architecture

Rather than relying on a single monolithic AI model, Zazalingo harnesses specialized Multi-LLM workflows and hybrid inference techniques across distinct engineering pipelines:

• Software & Systems Infrastructure: Advanced reasoning models (Claude / Opus architecture) utilized for system architecture design, API pipelines, and relational database bridging.
• Translation & Morphosyntactic Engine: High-throughput inference models (Gemini Flash series) operating under strict rule filters to enforce Zazaki orthography and ergative case structures when queries surpass Golden Corpus bounds.
• Data Mining & Synthetic Normalization: Analytical open-weight and proprietary models (DeepSeek and related architectures) deployed for corpus auditing, morphological extraction, and synthetic expansion testing.
• Verification & Guidance (RAG): Our proprietary local Golden Corpus acting as an authoritative arbiter, validating model outputs against authentic Zazaki linguistic rules.

Shared Responsibility for a Living Language

Zazalingo represents the resolve of an ancient oral tradition to articulate its own voice in the algorithmic era. Far beyond a few scripts or an interactive interface, this platform embodies nearly a year of arduous dedication, meticulous grammatical cataloging, and sleepless nights.

No computational system can encompass an entire language on its own; Zazalingo will mature only through your collective engagement, vigilance, and contributions. I warmly welcome reports of translation discrepancies, technical bugs, or constructive recommendations. Together, shoulder to shoulder, let us sustain and elevate our mother tongue in the digital universe.