Determinative Detection in Egyptian Hieroglyphic Sequences: A Sequence-Labelling Benchmark and Multi-Agent LLM Baseline

For over three thousand years, the ancient Egyptians wrote their history, their prayers, and their laws using a system of pictures that looks like a gallery of birds, legs, and geometric shapes to the modern eye. But this writing system was far more than just a list of sounds. It was a hybrid code where a single picture could do three different jobs at once: it could represent a sound, it could stand for a whole word, or it could act as a silent label. This third job, performed by a sign called a determinative, is the key to understanding the text. A determinative is a silent marker placed at the end of a word to tell the reader what kind of thing the word is. It might signal that the word refers to a person, an animal, a god, or a type of action. Because these signs are silent and often look like the other pictures in the sentence, figuring out which ones are just sounds and which ones are the silent labels is a difficult puzzle. Without solving this puzzle, computers cannot properly translate ancient Egyptian, because they cannot tell where one word ends and the next begins, nor can they understand the meaning of the words they find.

For decades, scholars have studied these silent labels by hand, but the sheer volume of text makes it impossible for humans to process everything alone. Recently, researchers have tried to teach computers to read these ancient scripts, but most of that work has focused on simply recognizing the pictures themselves, like identifying a drawing of a bird or a leg. Very little attention has been paid to the harder task of figuring out which pictures are acting as the silent labels. A new study from researchers in Brazil and Spain addresses this gap directly. They have created a computer system designed specifically to find these silent labels in sequences of ancient Egyptian writing. Instead of just guessing, the system uses a team of artificial intelligence agents that work together, checking the text from different angles to reach a conclusion that is both smart and reliable.

The researchers started by building a specialized library of knowledge. They gathered information on more than eighty different signs that can act as silent labels, organizing them into categories like people, animals, gods, and nature. They then created a test set of ninety-nine short sequences of ancient writing, which had been carefully checked and labeled by human experts. This set serves as a benchmark, a standard test to see how well a computer program can perform the task. The goal was to see if a machine could look at a string of symbols and correctly point out which ones were the silent labels and which ones were just sounds or parts of the word itself.

To solve this, the team built a system that acts like a small committee of experts. They used large language models, which are powerful computer programs trained on vast amounts of text, but they did not let them work alone. Instead, they set up two independent agents to analyze each sequence of writing. One agent looked at the meaning of the signs, asking questions like, "Does this picture fit the idea of a god or a person?" The other agent looked strictly at the structure and position of the signs, asking, "Is this sign at the end of the word, where silent labels usually go?" Crucially, these two agents did not talk to each other while they were working. They had to form their own opinions independently. This design ensures that if both agents agree on a label, it is a strong sign that they are right, rather than just copying each other's mistakes.

Once the two agents finished their analysis, a third agent acted as a synthesizer, bringing their findings together. It looked for where the two agreed and where they disagreed. Finally, a set of strict, unchangeable rules acted as a referee. This referee checked the suggestions against the knowledge base the researchers had built. It applied specific rules, such as "a sign meaning a vessel can only be a silent label if the text is about liquids or offerings." If the context did not fit, the rule rejected the suggestion, even if the artificial intelligence agents thought it was correct. This combination of flexible thinking from the AI and rigid rules from the referee created a system that could handle the complexity of the ancient language without making wild guesses.

The researchers tested their system using a rigorous method designed to ensure the results were not just a lucky fluke. They split their ninety-nine test sequences into five different groups. In each round of testing, they used four groups to train the system and the fifth group to test it, repeating this process until every single sequence had been tested exactly once. This approach gave them a clear picture of how the system would perform on new, unseen text. The results showed that the system was highly effective. It correctly identified the silent labels in the vast majority of cases, achieving a success rate of about eighty-two percent when looking at all types of labels equally, and about eighty-eight percent when looking at the total number of labels found. The consistency of these results across the different test groups showed that the system was robust and not dependent on a specific set of examples.

One of the most interesting findings was how the system handled difficult cases. The researchers discovered that the system worked best when the text was rich with information, containing three or more silent labels. In these complex sequences, the different clues from the meaning and the structure reinforced each other, making the correct answer obvious. However, the system struggled more with very short sequences that had no silent labels at all, or just one. In these sparse cases, the lack of surrounding clues made it harder for the system to be sure, leading to a few more errors. This suggests that the system is very good at finding patterns when there is plenty of evidence, but it needs more help when the text is very simple.

Another major innovation of this work is how it presents its answers. Instead of giving a single, final answer like a black box, the system provides a ranked list of possibilities. It offers a main interpretation, which is the most likely answer, but it also lists up to three alternative interpretations. For each alternative, the system explains exactly what extra evidence would be needed to make that alternative the correct one. For example, it might say, "We think this sign is a label for a god, but if the text were about a temple, it could also be a label for a building." This transparency turns the computer into a collaborative tool for human scholars. It does not replace the expert; instead, it lays out the options and the reasoning, allowing the human to make the final decision based on their deeper knowledge.

The study also highlighted specific challenges that remain. The system sometimes confused signs that have two very different jobs, such as a picture of walking legs that can mean "to walk" or act as a silent label for "action." Because the computer relies on context to decide which job the sign is doing, it occasionally made mistakes when the context was ambiguous. The researchers noted that these errors were not due to a flaw in the system's logic, but rather a limitation in the current knowledge base. By adding more examples and refining the rules for these tricky signs, the system could likely improve further.

Ultimately, this work establishes a new standard for how computers can interact with ancient writing systems. It proves that a hybrid approach, combining the flexible reasoning of modern artificial intelligence with the strict rules of human scholarship, can solve problems that neither could solve alone. The system does not just produce a translation; it produces a structured argument that can be checked, questioned, and improved. By making the reasoning visible and the alternatives clear, the researchers have created a tool that respects the complexity of the ancient language while opening the door for faster and more accurate analysis of the thousands of texts that remain to be read. This is not a magic solution that solves everything instantly, but it is a significant step forward, providing a reliable baseline for future research and a new way for machines and humans to work together to understand the past.