Can a lost script be read?

Some have been. The ones that remain unread are missing something specific, and in one case the argument is about whether it is writing at all.

ShallowsUncertainInspect the question

The question

Egyptian hieroglyphs fell to a bilingual inscription. Linear B fell to a guess about place names that pointed to Greek, a language its excavator had dismissed for forty years.

What remains unread is not harder in the same way. It is missing different things.

What decipherment needs

Known. Strongly supported by evidence and broadly established.

Strongly supported by evidence and broadly established.

What a decipherment needs

Two of three things make it possible: knowing the underlying language, having the same text in a known script, or having enough material for statistical attack.

Linear B had a large corpus and an unknown language, and was broken in 1952 by identifying the language as an early form of Greek. Egyptian had a bilingual text. Maya glyphs had a partial key in a colonial manuscript and a living related language.

Inspect

Why the Indus signs are different

The corpus is about four thousand inscriptions, the majority on stamp seals, averaging fewer than five signs. There is no bilingual text, the language is unknown, and the average inscription is shorter than this sentence.

In 2009 a study applied a statistical measure of how ordered the sign sequences are and reported that the Indus pattern matched known writing systems and stayed far from the non-linguistic sequences it was compared with, two of which were generated artificially.

The result was contested immediately and at length, on the grounds that the measure does not distinguish language from other structured symbol systems. Its authors replied that they had never claimed the measure proves language on its own, only that it adds weight to the script’s other language-like features. The exchange ran in a computational linguistics journal, and Sproat’s side of it is unusually direct about the reviewing he argues let the claim through.

It did not end there. In 2014 Sproat compared a larger set of writing and non-linguistic corpora and reported that the two measures he found useful class the Indus signs as non-linguistic; the original authors replied that they stand by their findings.

Two readings of the same signs

It encodes language

Uncertain. Meaningful scientific, historical or philosophical disagreement remains.

Sequences show the degree of order found in writing systems
Signs keep to preferred positions, as grammar would produce
Short texts are normal on seals in other literate cultures

It may not

Uncertain. Meaningful scientific, historical or philosophical disagreement remains.

Statistical order also appears in non-linguistic sign systems such as the deity symbols on Mesopotamian boundary stones
No long text on any medium, and none of the traces that lost manuscripts leave in other civilisations
Near Eastern emblem systems, religious and political, look much like this

Inspect

The average Indus inscription is under five signs long.

What is missing

Uncertain. Meaningful scientific, historical or philosophical disagreement remains.

Meaningful scientific, historical or philosophical disagreement remains.

What machine learning changes

Statistical and neural methods have been applied to lost scripts, and their clearest successes are re-runs. A 2019 model matched two thirds of Linear B’s cognates to their Greek counterparts, a decipherment people had made in 1952, and it was given Greek to work from.

Where the related language is unknown, the methods have little to hold. Tried on Iberian in 2021, one found no strong evidence for Basque as its relative, in line with most current scholarship. The bottleneck is material, not cleverness.

Inspect

What would break them

A bilingual inscription, or a long text, or a securely identified related language. All three are matters of excavation and luck rather than of method.

Linear A is the most promising of the group and has not yielded. Its signs are close enough to Linear B’s that the phonetic values of the later script can be read back onto the earlier one, which gives sounds without giving a language. That is what a related script buys on its own: you can roughly pronounce the words and still not know what most of them mean.