EJIE – Productivisation of AI-Based Language Tools
EJIE
Sector: Services
Business Case
How to run translation, transcription and synthesis models in service mode on day 1.
Objetivos
Technical: service scaling based on needs and autorecovery. Functional: When the translation can be improved, how do we detect and correct it?
Use case
The models were deployed in container mode on an on-premise solution (Red Hat openshift) with virtualised graphics resources (4gb slots of V100s cards). The models must “enter” the allocated graphical memory and the container must have “health” functions (liveness in Kubernetes language).
Infraestructura
On Premise
Tecnologías utilizadas
Machine learning and deep learning NLP
Datos utilizados
Translation: general language with media data (public) and administrative language with 100% IVAP translation service data. Transcription: base model + processed data (trials + eitb). Synthesis: 4 voices (3 eu + 1 es) with generated data + 2 inet data (fr, en).
Recursos utilizados
The team was formed by Vicomtech and EJIE’s Innovation group and the solution was deployed in the Basque Government’s data centres. It was necessary to hire for both data enrichment (transcription) and data generation from scratch (synthesis).
Dificultades y aprendizaje
On the technical level, it was necessary to internalise dynamics when working with graphic resources (GPU): acquisition, virtualisation, appropriate sizing of models to fixed-size slots, etc. Model improvement and correction are achieved through data, not programming. The generation of data is not an activity that can be contracted by EJIE, so this activity was carried out directly by the Basque Government.
KPIs (impacto en el negocio y métricas del modelo)
There are indicators of the number of requests made on a daily basis and segmented by group (device, ip, apikey, etc.). More than 200,000 requests per day by a collective of 30,000 unique individuals. The number of translation requests made to official services has decreased due to the assistance provided by language tools.
Financiación
This project was 100% executed with money from the budget with the participation of the language policy (HPS), the technology department (DTIC) and IVAP.
Colaboradores
Vicomtech was awarded the contract for the AI models. EJIE contributed more than 40 years of knowledge in IT management.
