Analyses have uncovered that DeepMind’s AI is significantly better at decoding the text than experts in human epigraphy are. This system could be used alongside humans as a collaborative tool to enhance historians’ understanding of a text and improve knowledge of ancient civilizations.
Ancient Texts are the Key to Understanding Past Civilizations
Ancient texts are the key to understanding the civilizations of the past. Scripts carved into stone, ceramic, and metal, as well as those written onto scrolls, have the potential to bring the thought processes, society, and history of past civilizations to life.
They are hugely valuable for understanding ancient cultural heritage. Fortunately, many thousands of writings have endured the hardships of time.
These texts are often incomplete due to the damage that has been inflicted on them over the hundreds of years that they have endured. This damage often eliminates sections of the prose which makes them difficult to translate.
Figuring out the text becomes akin to solving a puzzle. There is usually more than one single solution to how a text can be translated. The missing pieces make it difficult to determine which of the solutions fits best.
Trained epigraphists learn how to decode these writings by paying attention to grammatical and linguistic clues, considering the layout and shape of the text, investigating textual parallels, and examining the historical context.
The process is very time consuming and requires epigraphists to undergo lengthy studies to gain the required skills. It is also subject to human error due to interpretation.
To improve on the process of translating ancient texts, DeepMind, a world leader in Artificial Intelligence, built a machine learning system to do the work of an epigraphist, but with increased accuracy.
They developed PYTHIA, the world’s first ancient text restoration model. It was built with the capacity to use its neural networks to recover the missing characters from a damaged text and to give a more accurate translation of the original text.
The World’s First Machine Learning System
Researchers at DeepMind developed a machine learning system designed to manage long-term context information, in addition to dealing effectively with missing or broken characters and word representations.
They named the system Pythia, taking the name from the Greek priestess (also known as Oracle) who was believed to have the power to channel prophecies from the god Apollo himself.
To train Pythia, the team at DeepMind wrote a non-trivial pipeline to convert the PHI Greek inscriptions, the largest digital collection of ancient Greek inscriptions, into machine-actionable text, which they termed PHI-ML.
Sequences of damaged text were inputted into Pythia, which was trained to predict the character sequences, made up of hypothesized restorations of ancient Greek inscriptions. Pythia was trained to work on both the character and the word level, giving it the capability to deal with long-term context information, and manage incomplete word representations.
Pythia is More Accurate Than Humans at Decoding Text
DeepMind compared Pythia's accuracy in making predictions on PHI-ML against that of human epigraphists.
It was found that Pythia could achieve a 30.1% character error rate. For humans, this error rate was much greater at 57.3%. The ground-truth sequence was within the top 20 hypotheses made by Pythia in 73.5% of cases.
These figures demonstrate the power of Pythia as a system for ancient text restoration that could be effectively used alongside humans in the field of epigraphy.
More Applications on the Horizon
Pythia will profoundly impact the way historians work, giving them the ability to better understand previous civilizations by gaining access to more accurate interpretations of their writings.
The new system may cause a shift in the processes used in modern epigraphy, as the results of DeepMind’s study have shown Pythia’s value as a tool to be used alongside human epigraphists.
Pythia can assist in text restorations, providing multiple textual restorations along with the confidence level for each hypothesis. This is likely to change the way of working in the field.
The architecture of Pythia means that it is not only applicable to decoding ancient Greek, but it can also be used with other forms of ancient text, along with modern languages. This could see Pythia being used in numerous future applications.
Sources and Further Reading
Assael, Y., Sommerschield, T. and Prag, J. (2019). Restoring ancient text using deep learning: a case study on Greek epigraphy. [online] arXiv.org. Available at: https://arxiv.org/abs/1910.06262 [Accessed 17 Nov. 2019].