AI-based language technologies have demonstrated considerable success for civilian applications. However, the defence sector faces major bottlenecks preventing the development of these technologies. The confidentiality of the data (classified) coupled with the specificities of the defence environment (noise, phraseology, vocabulary…) represents a major challenge. Indeed, training AI models requires large annotated corpora specific to each application.

Generally classified, the preparation use of such data is subject to significant constraints, making the development of AI models highly expensive and very slow. Beyond, the resulting models are also classified and cannot be mutualised. ALADAN, a 42-month project, will design and develop a disruptive framework of development for AI-based language solutions for defence applications (i.e., spoken language identification, speech recognition, spoken term search, and text and speech translation), that will rely mainly on the use of non confidential data for training to target in-domain data using only a small amount of application-specific data for validation. It will enable the development of an in-domain solution within 6 months, with significantly reduced associated costs and delay.

ALADAN gathers 4 SMEs, i.e., Vocapia Research (coordinator), Crowdee, Lingea and IANUS Consulting, from 4 EU countries. ALADAN’s outcomes will be: i) language processing solutions covering multiple domains; ii) validation of the innovative development framework; iii) an easy-to-use toolkit permitting non-expert users to customize the models using classified data; and iv) a demonstrator including the targeted technologies with multidomain models for at least 2 defence use cases (defined with the MoDs). The open-domain development framework and the associated adaptation toolkit will be game changers in the adoption of language technologies by the forces.