Can Code Decipher the Indus Script? A Review of the Kee2u Project

What the project actually does

The Kee2u GitHub project, created by Keerthana Jayakumar with contributions from Soniaprogram, is an ambitious computational exploration. It begins with modern Tamil news text whose words, lemmas and parts of speech are known. The notebooks separate or curate morphemes, assign integer signs to lemmas and endings, spell some remaining material with CV and VC units, and assemble an artificial “logosyllabic Tamil” corpus. The project then compares aggregate patterns with published descriptions of Indus inscriptions.

This is a useful experiment in representation. It is not yet a reading of an Indus inscription. The transformation starts with known Tamil and deliberately makes it sign-like; it does not discover Tamil sounds or meanings from unknown Indus signs.

What our audit found

The committed machine-learning table contains 62,833 character positions, drawn from 9,581 token keys and 3,583 distinct surface forms. It has 24,521 positive morpheme-boundary labels and 38,312 non-boundaries. A saved support-vector-machine run reports 76.06% balanced accuracy, with F1 scores of 0.81 for non-boundaries and 0.71 for boundaries.

That result concerns segmentation within the project's modern-Tamil data. It does not test the language of Indus inscriptions. The random split is made after words have been expanded into character rows, so information from the same or repeated words can occur on both sides of the split. The character and prefix encoders are also fitted before splitting. A stronger test would hold out entire word types or documents and fit every transformation using training data alone.

A useful negative result

A separate experiment asks whether a sign's positional distribution can identify syllabic signs. In the saved test output, the minority “syllable” class has precision 0.31, recall 0.47 and F1 0.37; tuned neural models remain unstable. The repository itself concludes that position alone is insufficient. Publishing that failure is good research practice because it narrows what the statistic can support.

Similarity is not decipherment

Short-text statistics can reveal constraints. Frequency ranks, bigrams, positional preferences and entropy can help compare corpora. They cannot by themselves tell us what a sign sounded like or meant. The repository reports a modified power-law exponent of -1.13 for its constructed Tamil and cites -1.35 for Indus. Even if directly comparable, two nearby numbers do not distinguish common language ancestry from convergence, corpus design or the rules used to manufacture the Tamil encoding.

The average Indus text length, often given as about 4.5 signs, also cannot prove a logosyllabic system. Seal texts could contain names, titles, commodities, numerals, institutions or non-sentential formulae. Length does not reveal subject-object-verb grammar.

Why genetics cannot select the language

The Dravidian hypothesis is a serious scholarly proposal, but population ancestry and language are different variables. Ancient DNA can constrain migration and mixture; it cannot record what language a person spoke. Modern Tamil is also separated from the Mature Harappan period by millennia. A linguistic case needs reconstructed historical forms, regular correspondences, archaeological context and tests against other Dravidian and unrelated languages.

What would count as stronger evidence

A next-stage study should freeze its sign list, corpus, segmentation rules and success criteria before testing. It should deduplicate impressions of the same artefact; split by object, site and period; compare several reconstructed languages with known writing systems and non-linguistic controls; and publish confidence intervals and ablations. Most of all, any proposed sign values should make predictions on inscriptions withheld from every design decision.

Credits and sources

The review credits the repository authors for making their exploratory work public. It was audited at commit 73b0b9fe2acb1106f8b1b017c7350d35fa798997, dated 15 August 2021. The research paper cites the repository, Tamil Dependency Treebank, Rao and colleagues' entropy study and reply, Sproat's critique, Yadav and colleagues' n-gram study, Mahadevan's concordance, Parpola's monograph, ancient-DNA studies, and recent surveys of machine learning for ancient languages.