A Gentle Introduction to Natural Language Processing (NLP)
When a computer reads your Urdu voice note, filters your spam, or translates a message for a cousin abroad, that is natural language processing — NLP, the field teaching machines human language. In English it works so well it has become boring. In our languages it is a harder, far more interesting story.
What NLP actually does
The field splits into two directions. Understanding means pulling meaning out of text: is this review angry, is this email spam, who does "she" refer to in this sentence. Generation means producing sensible language: a reply, a summary, a translation. Your day already touches both — Gmail's spam filter, Google Translate, YouTube's auto-captions, the chatbot on a bank's website. The tasks have barely changed in thirty years; the machinery has changed completely, and large language models are now most of it.
Why Urdu is a different game
NLP is trained on text, and the internet's text is overwhelmingly English. Urdu's slice is small, and the spoken reality is messier still: most Pakistanis write to each other in Roman Urdu sprinkled with English, a mixture no tidy grammar captures. Our scripts are rich — Urdu's Nastaliq alone defeats plenty of rendering systems. Tokenizers, the tools that chop text into pieces a model can process, split Urdu into more fragments than English, which means longer prompts, higher costs, and more places to go wrong. A model that excels in English can still mangle a Sindhi sentence it has barely ever seen.
The gap is closing
This is the hopeful part. Meta's NLLB-200 project pushed machine translation into two hundred languages, including Sindhi and Pashto — the kind the industry long ignored — and SeamlessM4T extended the work to real-time speech translation. Pakistan's own reported effort, Indus-1, a language model fine-tuned for Urdu, Punjabi, Sindhi, and Pashto and released with open weights by a university consortium, exists precisely because nobody else was going to prioritise us. The practical uses are unglamorous and real: bank chatbots that answer in Urdu, school lessons explained in Punjabi, government forms that stop assuming English fluency.
Why it matters beyond convenience
A language that machines cannot process is a language locked out of the next generation of services — search, education, commerce. In a country where English is the language of the elite and everything important happens in everything else, decent NLP in our languages is not a luxury. It decides whether the next wave of tools serves Pakistanis or skips them.
Some words need no model to process. "Sumud" — staying, holding on — carries Palestine's whole story in five letters, and it is the quietest, strongest sentence in this post. If you understand it, you already understand more than a machine can.
Meanwhile, the best language processing we know is still human, and it works at our desk: HTG Travels in Sialkot handles Urdu, Punjabi, and English daily — visa cover letters, Umrah group instructions, translated document lists for parents travelling alone. Machines translate; we make sure the meaning survives the embassy.
