Telegraph English: Semantic Prompt Compression via Structured Symbolic Rewriting
Telegraph English (TE) is a novel prompt-compression protocol that rewrites natural language into a symbol-rich, structured dialect of atomic fact lines, achieving superior accuracy and semantic indexing capabilities across various models compared to existing token-deletion methods like LLMLingua-2.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you have a very long, wordy letter from a friend. You need to send it to someone else, but you want to save on postage (tokens) and make sure the recipient understands the most important parts without getting lost in the fluff.
Currently, the standard way to do this (called LLMLingua-2) is like a game of "Redaction." You take a red pen and cross out 50% of the words, hoping the remaining sentences still make sense. The problem? You might accidentally cross out the word "because" or "therefore," leaving the reader to guess how the ideas connect. Or you might cut out a specific number that turns out to be the most important part of the story.
This paper introduces a new method called Telegraph English (TE). Instead of just crossing out words, TE acts like a master translator who rewrites your entire letter into a super-efficient, code-like language.
Here is how it works, using simple analogies:
1. The "Redaction" vs. The "Rewrite"
- The Old Way (Token Deletion): Imagine you are editing a novel by cutting out every other word. You might end up with: "The cat sat on the... rug." It's shorter, but the flow is broken. If you cut out "not," the meaning flips entirely.
- The New Way (Telegraph English): Imagine taking that same novel and rewriting it as a series of bullet points using a secret shorthand.
- Original: "According to research by Johnson and colleagues (2023), the application of machine learning techniques to medical diagnostics resulted in a 27.5% increase in early detection rates while simultaneously reducing false positives by approximately 12% compared to traditional methods."
- Telegraph English:
ML→MEDICAL-DIAGNOSTICS: EARLY-DETECTION+27.5% ∧FALSE-POSITIVE-12% [JOHNSON:2023] - The Result: The sentence is cut down from 68 words to 14 "tokens," but every single fact, number, and relationship is still there, clearly labeled.
2. The "Lego Brick" Advantage
The paper claims that TE has a special superpower: Compression and Organization happen at the same time.
- The Old Way: When you redact a text, you get a smaller pile of messy words. If you want to find a specific fact later, you have to read through the whole messy pile again.
- The New Way: TE breaks the text into "Atomic Fact Lines." Think of this like taking a big, messy block of clay and turning it into individual, labeled Lego bricks.
- Each line is one single fact.
- Each line is tagged (e.g.,
PAST:,LIKELY:,CITATION:). - Because every line is a self-contained fact, you can instantly pull out just the "dates" or just the "numbers" without re-reading the whole document. It's like having a library where every book is already sorted by its topic on the spine.
3. Why It Works Better for "Smarter" and "Dumber" Computers
The researchers tested this on five different AI models (from very smart ones to smaller, cheaper ones).
- For the Smart Models: TE performed just as well as the old method, but with half the cost.
- For the Smaller Models: This is where TE shines. Smaller AI models sometimes struggle to "fill in the blanks" when words are missing.
- Analogy: If you give a small child a puzzle with half the pieces missing (Token Deletion), they might guess the picture wrong. If you give them a puzzle where every piece is clearly labeled and connected (Telegraph English), they can solve it perfectly.
- The paper found that on tricky, detail-heavy questions, TE helped the smaller models get the right answer up to 11% more often than the old method.
4. The "One-Time Rewrite, Forever Manageable" Rule
The paper introduces a concept called "Compress Once, Manage Continuously."
- The Old Way: You compress the text once to send it. If the AI generates a long answer, that answer comes back as a giant, uncompressed wall of text. If you need to send that answer to a third person, you have to compress it all over again.
- The New Way: Because TE turns text into structured "fact bricks," the AI's answer can also be written in TE. This means the whole conversation stays compressed and organized. You don't have to re-compress anything; the structure is built-in.
The Catch (Limitations)
The paper is honest about the downsides:
- It needs a "Translator": To turn your text into Telegraph English, you need to run an AI model first. This takes time and money upfront. If you are sending a message that will only be read once and never used again, this extra step might not be worth it.
- English Only: This specific "shorthand" is designed for English. It wouldn't work immediately for languages with very different structures (like Chinese or Arabic) without a major redesign.
- Strict Rules: The system is very rigid. It forces facts into specific formats. If the input is already very short and dense, the system might actually make it longer because it refuses to delete any information (fidelity is more important than brevity).
The Bottom Line
Telegraph English isn't just about making text shorter; it's about making text smarter. It trades the messy, natural flow of human language for a clean, structured, symbol-rich format that is easier for computers to read, store, and remember. It turns a "degraded copy" of a document into a "structured database" of facts.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.