The Kashmiri language is spoken by over 7 million people, yet it remains one of the most underrepresented languages in modern AI systems. There are no dedicated machine translation APIs, no large-scale parallel corpora, and only a handful of academic papers addressing it. This is the problem we set out to solve — and this is how we built the first dedicated NLP and human evaluation platform for the Kashmiri language.
Why Kashmiri Needs Its Own NLP Research
Kashmiri (کٲشُر / कॉशुर) is an Indo-Aryan language with a unique phonological system, complex morphology, and a dual writing tradition — Nastaliq script (right-to-left) and Devanagari (left-to-right). These characteristics make it particularly challenging for general-purpose translation systems, which are primarily trained on high-resource languages.
Current MT systems produce poor-quality translations for Kashmiri — often missing idiomatic expressions, code-switched phrases with Urdu or Hindi, and culturally specific terminology. The root cause is data scarcity: there is no large-scale, publicly available Kashmiri→English parallel corpus.
What We Built
KashmirAI Research is a two-part system: a machine learning pipeline for training and evaluating translation models, and a web-based human evaluation platform where native Kashmiri speakers assess output quality.
🏗️ Platform Components
- Parallel Corpus Construction — A curated Kashmiri→English sentence dataset built from diverse domains
- LLM Fine-tuning Pipeline — Fine-tuning open-source models on Kaggle GPU infrastructure using LoRA
- Human Evaluation System — Real-time web app for pairwise preference judgments by native speakers
- Annotator Reliability Dashboard — Cohen's Kappa, trust scores, and inter-annotator agreement metrics
Low-Resource NLP: The Core Challenge
Low-resource NLP refers to building language technologies for languages with limited training data. Unlike English, which has billions of training sentences, Kashmiri has only thousands of publicly available parallel sentences. This means we cannot simply pre-train a transformer from scratch — we must leverage transfer learning and fine-tuning techniques.
Our approach uses parameter-efficient fine-tuning (PEFT) — specifically LoRA (Low-Rank Adaptation) — to adapt large pre-trained multilingual models to the Kashmiri→English translation task without catastrophic forgetting.
Our Research Methodology
- Pairwise Comparison — Evaluators choose the better translation out of two anonymized systems, reducing scoring bias.
- MQM Error Tagging — Evaluators tag specific error spans (fluency, adequacy, terminology) rather than assigning a single score.
- Control Sentences — Hidden sentences with known correct translations are interspersed to detect unreliable annotators.
- Inter-Annotator Agreement — Cohen's Kappa is computed across all overlapping judgments to measure evaluation reliability.
Future Roadmap
Native Kashmiri Speaker?
We need your help evaluating translation quality. Each session takes 10–15 minutes and directly advances AI for the Kashmiri language.
Start Evaluating →