After several months of tests and different types of training, the first massive Artificial Intelligence system in the Spanish language is capable of generating its own texts and summarising existing ones. MarIA is a project that has been promoted by the Secretary of State for Digitalisation and Artificial Intelligence and developed by the National Supercomputing Centre, based on the web archives of the National Library of Spain (BNE).

This is a very important step forward in this field, as it is the first artificial intelligence system expert in understanding and writing in Spanish. As part of the Language Technology Plan, this tool aims to contribute to the development of a digital economy in Spanish, thanks to the potential that developers can find in it.

The challenge of creating the language assistants of the future

MarIA-style language models are the cornerstone of the development of the natural language processing, machine translation and conversational systems that are so necessary to understand and automatically replicate language. MarIA is an artificial intelligence system made up of deep neural networks that have been trained to acquire an understanding of the language, its lexicon and its mechanisms for expressing meaning and writing at an expert level.

Thanks to this groundwork, developers can create language-related tools capable of classifying documents, making corrections or developing translation tools.

The first version of MarIA was developed with RoBERTa, a technology that creates language models of the "encoder" type, capable of generating an interpretation that can be used to categorise documents, find semantic similarities in different texts or detect the sentiments expressed in them.

Thus, the latest version of MarIA has been developed with GPT-2, a more advanced technology that creates generative decoder models and adds features to the system. Thanks to these decoder models, the latest version of MarIA is able to generate new text from a previous example, which is very useful for summarising, simplifying large amounts of information, generating questions and answers and even holding a dialogue.

Advances such as the above make MarIA a tool that, with training adapted to specific tasks, can be of great use to developers, companies and public administrations. Along these lines, similar models that have been developed in English are used to generate text suggestions in writing applications, summarise contracts or search for specific information in large text databases in order to subsequently relate it to other relevant information.

In other words, in addition to writing texts from headlines or words, MarIA can understand not only abstract concepts, but also their context.

More than 135 billion words at the service of artificial intelligence

To be precise, MarIA has been trained with 135,733,450,668 words from millions of web pages collected by the National Library, which occupy a total of 570 Gigabytes of information. The MareNostrum supercomputer at the National Supercomputing Centre in Barcelona was used for the training, and a computing power of 9.7 trillion operations (969 exaflops) was required.

Bearing in mind that one of the first steps in designing a language model is to build a corpus of words and phrases that serves as a database to train the system itself, in the case of MarIA, it was necessary to carry out a screening to eliminate all the fragments of text that were not "well-formed language" (numerical elements, graphics, sentences that do not end, erroneous encodings, etc.) and thus train the AI correctly.

Due to the volume of information it handles, MarIA is already the third largest artificial intelligence system for understanding and writing with the largest number of massive open-access models. Only the language models developed for English and Mandarin are ahead of it. This has been possible mainly for two reasons. On the one hand, due to the high level of digitisation of the National Library's heritage and, on the other hand, thanks to the existence of a National Supercomputing Centre with supercomputers such as the MareNostrum 4.

The role of BNE datasets

Since it launched its own open data portal (datos.bne.es) in 2014, the BNE has been committed to bringing the data available to it and in its custody closer: data on the works it preserves, but also on authors, controlled vocabularies of subjects and geographical terms, among others.

In recent years, the educational platform BNEscolar has also been developed, which seeks to offer digital content from the Hispánica Digital Library's documentary collection that may be of interest to the educational community.

Likewise, and in order to comply with international standards of description and interoperability, the BNE data are identified by means of URIs and linked conceptual models, through semantic technologies and offered in open and reusable formats. In addition, they have a high level of standardisation.

Next steps

Thus, and with the aim of perfecting and expanding the possibilities of use of MarIA, it is intended that the current version will give way to others specialised in more specific areas of knowledge. Given that it is an artificial intelligence system dedicated to understanding and generating text, it is essential for it to be able to cope with lexicons and specialised sets of information.

To this end, the PlanTL will continue to expand MarIA to adapt to new technological developments in natural language processing (more complex models than the GPT-2 now implemented, trained with larger amounts of data) and will seek ways to create workspaces to facilitate the use of MarIA by companies and research groups.

Content prepared by the datos.gob.es team.