From Expression to Action in Intelligent Systems
Arabic Is Not a Translation Layer
Fluency is the surface. Fidelity means that intent survives representation, inference, and tool choice all the way to the right action.
The decision brief“When language becomes an interface to action, a language error becomes a systems risk.”
The decision brief
A practical starting point
Evaluate the action, not only the Arabic response: test invariant intent, consequential differences, and the decision to clarify.
- Distinguish linguistic quality from task performance.
- Identify how dialects, context and speech affect service design and evaluation.
- Frame comparisons around users and use cases, rather than a single model ranking.
Central question
Same intent. Will the machine take the same action?
The opening scene
Five expressions, one intent
Five messages reach a card-support system. One uses formal Arabic; the others use everyday Gulf, Egyptian, Levantine, and code-switched Arabic. Every customer wants the same outcome: freeze the card for now, but do not cancel it.
One message reads: «أبغى أوقف البطاقة شوي، مو ألغيها»—roughly, “I want to stop the card for a while, not cancel it.” Another switches directly to the English words “freeze” and “cancel.” Will the system preserve the distinction and take the same action?
The answer is decided below the surface. Encoding, tokenization, embeddings, training distribution, and context shape the representation before meaning reaches a tool, field value, or banking operation.
The lesson reaches beyond one language: fluency can conceal operational failure. Arabic makes the problem vivid because form varies widely while a small change in negation, person, amount, or time can reverse the required action.
Inside language representation
What a person sees as a sentence, a model sees as a sequence of units. Choosing those units is an engineering decision that affects sequence length, effective context, cost, latency, and what can be learned efficiently. Tokenization does not explain every quality difference, but it makes a useful truth visible: a quiet decision below the surface can reshape experience above it.
Arabic brings the reality of attached clitics, rich morphology, absent diacritics, dialect variation, and code-switching. The goal is not to romanticize complexity or turn it into an excuse. It is to build evaluations that reflect the language people actually use while trying to complete consequential tasks.
From response to action
In a conversational system, misunderstanding may produce an odd reply. In an agentic system, it may cancel a card instead of freezing it, select the wrong beneficiary, alter a date, or call the wrong tool. Style evaluation is no longer enough. Entities, negation, time, and scope must be traced through to the executed operation.
Sometimes the system must resist the urge to answer. If ambiguity can change the action and the cost of asking is small compared with the cost of error, “Do you mean a temporary freeze or permanent cancellation?” is a sign of intelligence. Clarification is a safety decision, not a failure of fluency.
Intent-to-Action Fidelity
The talk proposes Intent-to-Action Fidelity as an evaluation lens. We need invariance to irrelevant variation: different phrasings and dialects expressing the same intent should reach the same action. We also need sensitivity to consequential variation: a change in negation, amount, person, time, or scope must change the result.
The framework connects layers that are often evaluated separately. Fluency belongs to the visible text; fidelity runs through representation, interpretation, tool and parameter selection, confirmation, and execution. The useful test compares actions across varied expressions and treats clarification as the correct outcome when meaning is insufficient.
Evaluate the action
The frontier is moving from text quality toward task completion: dialects, speech, code-switching, Arabic retrieval, specialized and multilingual models, tool use, and voice agents. Models will change. The durable question is whether intent survived the entire computational route.
The final action is to assemble real expressions for a consequential task and test the outcome, not only the wording. Hold constant what should not change, vary what must alter the action, and include cases where the correct behavior is to ask before execution.
Explore the idea
The intent-fidelity test
Carry one intent through linguistic variation to the final action.
Phrase
Which forms, dialects, and code-switched patterns do people actually use?
All perspectives
Phrase
Which forms, dialects, and code-switched patterns do people actually use?
Context
Which entities, times, negations, and scope distinctions change meaning?
Intent
What should remain invariant, and which difference must change the result?
Action
Was the right operation executed, or should the system have clarified first?
An illustrative scene to explore
Windows into the idea
Questions that open another door
Is fluency evidence of understanding?
It tells us that the text sounds natural. Task understanding appears when the system preserves negation, time, entities, and scope, then selects the correct tool and action.
Is a specialized Arabic model always better?
It depends on the user, task, domain, cost, latency, and evaluation. A specialist may excel in one setting; a multilingual model may benefit from knowledge transfer across languages.
Why should a non-specialist care about tokenization?
It is a silent decision that reshapes text before the model sees it and can affect request length, context, cost, and representation. It shows how a small technical choice can alter a large public experience.
Are dialects noise to be normalized away?
Dialects are a fact of use. A capable system distinguishes changes in form from changes in meaning instead of demanding that people speak the language of the laboratory.
When is a question better than an answer?
When ambiguity can change the action. If asking is cheap and error is costly, clarification is part of intelligence and safety.
The idea’s bookshelf
Works that illuminate the subject and offer paths for further reading.
- Introduction to Arabic Natural Language Processing
The publisher page for the 2010 reference covering script, phonology, morphology, syntax, semantics, and notes on dialects in Arabic NLP.
- SentencePiece: A Simple and Language Independent Subword Tokenizer and Detokenizer for Neural Text Processing
An ACL Anthology proceedings paper from EMNLP 2018 describing a tool that learns subword units directly from raw sentences.
Who it is designed for
- Arabic AI, product and user-experience teams
- Researchers in language processing, speech and evaluation
- Organisations building Arabic or multilingual services
When it fits
For Arabic service development, model selection or evaluation design for Arabic-speaking users.
What it does not promise
Not a promotion for a particular model or a claim that specialised models always outperform multilingual ones.
One central idea. Depth to match the room.
These formats are proposals. Readiness, duration and interaction are confirmed for each invitation. These are the languages of the website descriptions. Live delivery languages and interpretation arrangements are confirmed separately for each engagement.