📄 LLMs as Compiler for Arabic Programming Language

Summary

The paper addresses the fact that nearly all mainstream programming languages use English keywords. This is a barrier for Arabic speakers. Earlier Arabic languages such as Sakhr Basic, ARLOGO and Phoenix each needed a hand-built compiler. The authors propose APL (Arabic Programming Language). It has no conventional compiler. Instead, GPT-4 acts as a "semi-compiler": it rewrites APL source into Python, and PyRunner then executes the result. APL is written right-to-left in Arabic script, marks blocks with curly braces instead of indentation, and puts one statement per line. A single engineered prompt turns Arabic keywords into Python ones and gives identifiers English names. It also leaves Arabic string literals unchanged and reports syntax errors. The release has three parts: a Python library on GitHub built around a Compiler class, a two-pane web playground, and the paper itself. The paper reports no quantitative evaluation, only worked examples. Its value is mainly as a proof of concept: a prompted LLM can stand in for the lexer, parser and code generator of a small new language.

Key points

• APL supports only six constructs: variables, operations, reading .txt files, if-conditions, loops and functions. The authors kept it small on purpose because chained LLM prompts are fragile.
• The source code follows three surface rules: Arabic script, {} block delimiters instead of Python indentation, and one operation per line with no ;.
• Translation uses GPT-4 with one system-style prompt that:
  • maps Arabic keywords to Python;
  • renames variables into English;
  • checks syntax and returns a specific error message;
  • outputs plain text rather than markdown-fenced code.
• The tooling consists of:
  • a GitHub Python library whose Compiler class is initialised with an API key;
  • a playground with an Arabic right-to-left input pane on the right and the generated Python on the left.
• The authors name two main challenges:
  • separating Arabic text that must be translated (keywords, names) from Arabic text that must be kept (string contents);
  • the deployment cost of the GPT-4 API.
• As future work they plan to fine-tune an open-source model.

Interesting ideas

• I1. The input syntax was changed to suit the LLM rather than humans: explicit braces replace indentation, because copy-paste damages whitespace and the model handles visible delimiters more reliably. — For Mova (learn): When sending sentences or trees to an LLM teacher tier, serialize structure with explicit brackets or IDs instead of relying on layout, so whitespace damage cannot corrupt silver annotations.
• I2. The language is deliberately cut down to a few constructs, because each additional feature makes a chained LLM pipeline more fragile; scope is traded for reliability. — For Mova (learn): Give LLM teachers narrow, closed-inventory tasks (e.g., one frame type or one relation at a time) rather than open-ended full annotation, then combine the results deterministically.
• I3. The core difficulty is a split between two kinds of tokens: structural Arabic text must be translated, while Arabic inside string literals must pass through untouched. — For Mova (tokens): The tokenizer could tag spans such as quotations, code and names as "opaque content" early, so later normalisation, lemmatisation or translation steps leave them verbatim.
• I4. The translator does more than convert code: it also detects syntax errors and returns a targeted message telling the user what to fix, turning failure into feedback. — For Mova (answer): When a question or word problem cannot be mapped into the text-world, Mova could report which step failed and why instead of giving a silent wrong answer, in line with its explainability goal.
• I5. The playground shows the source and the generated output side by side so users can compare them line by line and debug easily. — For Mova (back): The text → graph → text compressor could show the original and the regenerated sentence in aligned panes, making meaning loss in the round trip easy to spot.