Computational analysis of symbols used 4,000 years ago by a long-lost Indus Valley civilization suggests they represent a spoken language. Some frustrated linguists thought the symbols were merely pretty pictures.
"The underlying grammatical structure seems similar to what's found in many languages," said University of Washington computer scientist Rajesh Rao.
The Indus script, used between 2,600 and 1,900 B.C. in what is now eastern Pakistan and northwest India, belonged to a civilization as sophisticated as its Mesopotamian and Egyptian contemporaries. However, it left fewer linguistic remains. Archaeologists have uncovered about 1,500 unique inscriptions from fragments of pottery, tablets and seals. The longest inscription is just 27 signs long.
In 1877, British archaeologist Alexander Cunningham hypothesized that the Indus script was a forerunner of modern-day Brahmic scripts, used from Central to Southeast Asia. Other researchers disagreed. Fueled by scores of competing and ultimately unsuccessful attempts to decipher the script, that contentious state of affairs has persisted to the present.
Among the languages linked to the mysterious script are Chinese Lolo, Sumerian, Egyptian, Dravidian, Indo-Aryan, Old Slavic, even Easter Island — and, finally, no language at all. In 2004, linguist Steve Farmer published a paper asserting that the Indus script was nothing more than political and religious symbols. It was a controversial notion, but not an unpopular one.
Rao, a machine learning specialist who read about the Indus script in high school and decided to apply his expertise to the script while on sabbatical in Inda, may have solved the language-versus-symbol question, if not the script itself.
"One of the main questions in machine learning is how to generalize rules from a limited amount of data," said Rao. "Even though we can't read it, we can look at the patterns and get the underlying grammatical structure."
Rao's team used pattern-analyzing software running what's known as a Markov model, a computational tool used to map system dynamics.
They fed the program sequences of four spoken languages: ancient Sumerian, Sanskrit and Old Tamil, as well as modern English. Then they gave it samples of four non-spoken communication systems: human DNA, Fortran, bacterial protein sequences and an artificial language.
The program calculated the level of order present in each language. Non-spoken languages were either highly ordered, with symbols and structures following each other in unvarying ways, or utterly chaotic. Spoken languages fell in the middle.
When they seeded the program with fragments of Indus script, it returned with grammatical rules based on patterns of symbol arrangement. These proved to be moderately ordered, just like spoken languages.
As for the meaning of the script, the program remained silent.
"It's a useful paper," said University of Helsinki archaeologist Asko Parpola, an authority on Indus scripts, "but it doesn't really further our understanding of the script."
Parpola said the primary obstacle confronting decipherers of fragmentary Indus scripts — the difficulty of testing their hypotheses — remains unchanged.
But according to Rao, this early analysis provides a foundation for a more comprehensive understanding of Indus script grammar, and ultimately its meaning.
"The next step is to create a grammar from the data that we have," he said. "Then we can ask, is this grammar similar to those of the Sanskrit or Indo-European or Dravidian languages? This will give us a language to compare it to."
"It's only recently that archaeologists have started to apply computational approaches in a rigid manner," said Rao. "The time is ripe."