Google DeepMind Aeneas: AI for Contextualizing Ancient Inscriptions

Google DeepMind has introduced Aeneas, a multimodal generative neural network designed to help historians interpret, attribute, and restore fragmentary Latin inscriptions. By automating the identification of textual and contextual parallels, Aeneas accelerates the process of dating and placing ancient texts that are often weathered or damaged.

Advanced Capabilities for Epigraphic Analysis

Aeneas provides several key technical advancements over previous models, including its predecessor Ithaca (which focused on ancient Greek inscriptions). Its primary capabilities include:

  • Parallels Search: The model creates "historical fingerprints" (embeddings) for each text to identify deep connections across a vast collection of Latin inscriptions, allowing historians to situate fragments within a broader context.
  • Multimodal Provenance Determination: Aeneas is the first model to determine geographical provenance by analyzing both textual data and visual information, such as images of the inscriptions.
  • Flexible Text Restoration: The model can restore gaps in texts where the missing length is unknown, a significant improvement for dealing with heavily damaged materials.
  • State-of-the-Art Performance: Aeneas establishes new benchmarks for predicting the date and location of inscriptions and restoring damaged text.

Technical Architecture and Training

Aeneas is a multimodal generative neural network that processes text and image inputs using a transformer-based decoder.

Data Sources and the Latin Epigraphic Dataset (LED)

To train the model, researchers curated the Latin Epigraphic Dataset (LED), which harmonizes over 176,000 Latin inscriptions from three major digital collections:

  1. The Epigraphic Database Roma (EDR)
  2. The Epigraphic Database Heidelberg (EDH)
  3. The Epigraphic Database Clauss Slaby (EDCS-ELT)

Model Mechanism

The model utilizes specialized networks for character restoration and dating based on text, while geographical attribution incorporates image data. The contextualization mechanism uses embeddings to encode language, content, provenance, and relationships between inscriptions, allowing the decoder to retrieve and rank similar parallels from the LED.

Performance Benchmarks and Validation

Aeneas demonstrates superior performance in chronological grouping compared to general-purpose Latin LLMs. Key metrics include:

  • Restoration Accuracy: A Top-20 accuracy of 73% for gaps up to ten characters, and 58% accuracy when the restoration length is unknown.
  • Geographical Attribution: 72% accuracy in attributing an inscription to one of 62 ancient Roman provinces using visual data.
  • Dating Precision: The model typically places a text within 13 years of the date ranges provided by historians.
  • Interpretability: The model provides saliency maps to highlight which specific inputs influenced its predictions.

Application to Historical Debates: The Res Gestae

To test the model on a real-world research debate, Aeneas was applied to the Res Gestae Divi Augusti (Emperor Augustus' account of his achievements). Rather than providing a single date, Aeneas produced a probabilistic distribution with two peaks: a smaller peak around 10-1 BCE and a larger peak between 10-20 CE. This quantitative approach captured prevailing scholarly hypotheses based on subtle linguistic features and official titles mentioned in the text.

Collaborative Impact and Accessibility

In a study involving 23 historians, the most effective results were achieved when experts used Aeneas' contextual parallels alongside its restoration and attribution predictions. Historians reported that the tool accelerated their work and helped identify new parallels that changed their perception of specific inscriptions.

Availability and Educational Integration

Google DeepMind has made Aeneas available through the following channels:

  • Interactive Tool: Freely available at predictingthepast.com.
  • Open Source: The code and dataset are available on GitHub.
  • Ithaca Upgrade: The ancient Greek model, Ithaca, has been been upgraded to be powered by Aeneas, adding contextualization functions and improved restoration capabilities.
  • Education: A new teaching syllabus has been co-designed to bridge technical AI skills with historical thinking, aligning with frameworks from the European Commission, UNESCO, and the OECD.

Sources