Studying the laws: this is one of the tasks that Minerva is facing, the first chatbot in the world to speak a national language born at the Sapienza University of Rome thanks to the Natural Language Processing group (NLP) directed by Roberto Navigli, professor of artificial intelligence at Sapienza. Navigli himself tells Ansa, among the winners of the Capo d’Orlando prize with the Nobel John Jumper of the Google Deepmind and with the Christian paleontologist from the stone of the Museum of Natural History of Milan.
Released in November 2024 and developed in the context of the PNRR Fair project (Future Ai Research), Minerva is the first of the great linguistic models (LLM) to work in the Italian language: a peculiarity that is inspiring other research groups in Europe, such as the German one of the University of Hanover. “Minerva is interesting because the language of the country in which it is born, at the same level as the English language: no other model has these characteristics,” observes Navigli. It was probably this peculiarity that aroused the interest of the public administration. Minerva thus began to “study” the laws, but it is only the first response to the numerous requests that are also coming from other areas, such as health (for example with the reading of medical records), culture and education. These customizations are developed jointly with the Babelscape university spinoff.
Talking Italian offers many advantages, but there are gaps to fill. “The first – Navigli notes – is the lack of data. Minerva in fact reads the web, but the web in Italian is smaller”.
The first version released in November 2024 was in fact trained with about 1,000 billion words in Italian and as many in English, in all 7.4 billion parameters. “Today it is considered a small model”. For this reason Navigli and his group are studying the version 2 of Minerva: “It will be a refinement of the version of November 2024 and we are using a technique to expand the input to tens of thousands of words, almost the size of a book, and we plan to release version 2 with multiple parameters, more examples and new data”.
Now, he continues, “Minerva is able to read books and his ability to understand his ability to understand now extends to the analysis and the answer to questions on texts of considerable length. One of the next objectives is the insertion of the research on the web at 360 degrees, being able to converse on any topic. In the end of 2025 we aim for multimodality, that is, the ability to understand texts, images and perhaps even audio, of reasoning “.
Reserved reproduction © Copyright Ansa
Source: Ansa
I am a journalist with over 6 years of experience working in the news industry. I currently work as an author for Global Happenings, and my coverage focuses on Technology news. I have written for various publications, including Reuters, The New York Times and The Guardian.
