How to Brief a Natural Language Processing (NLP) Project, Including Arabic, So the Right Engineer Understands It

Natural language processing (NLP) is the part of AI that works with text and speech: understanding customer messages, classifying them, extracting information, summarising and answering them. NLP projects in Arabic have their own quirks, and a vague post gets proposals so different they can’t be compared. This guide helps you describe your project in the AI & Machine Learning category on Farasa. For the general rules, read our guide to posting software projects.

“AI” isn’t enough. Say which task you want: a chatbot that answers your customers’ questions; classifying messages (complaint, enquiry, request); extracting information from text (names, dates, amounts); summarising long documents; sentiment analysis of comments; smart search across your document archive; or translation. Each task has a different approach, cost and data need.

State clearly which language will be processed: Modern Standard Arabic, Iraqi dialect, a mix of Arabic and English, or Arabic typed in Latin letters online. A model that does well on formal news can stumble on WhatsApp messages in a local dialect. Attach dozens of real examples if you can; they describe your problem better than anything else.

Say how many messages or documents you have, where they come from, and whether they are labelled (has someone tagged each message by hand). Some projects need no large training set because ready-made language models can do the task with the right instructions; others need thousands of labelled examples. A good applicant will ask and recommend what fits, and it helps if you show what you have from the start.

Agree a measure before starting: “classifies 90% of messages correctly on a sample of 200 messages the model has not seen,” or “answers common questions without inventing information.” Ask for testing on your real text, not on ready-made examples. And think about the cost of errors: a wrong complaint tag is different from a wrong legal or medical answer.

Many projects today are solved with ready-made language models through an API, which is faster and cheaper but means sending your text to an outside service. If your data is sensitive or heavy usage cost matters, you may need a model that runs on your own server. State your privacy and cost constraints and ask applicants what they would choose and why.

Where will it run? Inside your website, in WhatsApp or Telegram, or as a service your own system calls? Do you need a dashboard to follow conversations? Does the customer get handed to a person when the system can’t cope? These details decide the size of the job more than the model itself.

Text may contain names, phone numbers and financial details. Don’t send sensitive data before agreeing terms, and begin with a sample with personal details masked. Say whether the data must stay in Iraq or on your own server.

“A system that sorts the customer messages of a delivery company in Baghdad into: enquiry, complaint, address change, cancellation. Messages are in Iraqi dialect and mixed Arabic and English; I have 5,000 messages, 800 of them labelled by hand. I want at least 90% accuracy on a fresh test sample, with anything the system doesn’t understand passed to a staff member. It will run as an API that my order system calls. Budget: fixed price for the first phase (classification) within a range (put your range here). Start your proposal by naming an NLP project you’ve delivered and how you evaluated it.”

Name the task, describe your language and dialect, show your data, agree a success measure tested on your own text, and state your privacy and integration limits. That way you get an engineer who understands your real problem. Post your NLP project on Farasa.