We knew it better under the name Pôle Emploi. Renamed France Travail, the public administrative establishment responsible for employment in France has not changed its mission. However, it seems that the organization is turning a corner with the gradual adoption over several years of new technologies, notably artificial intelligence, as stated by Agathe Ravilly, product manager at France Travail: “We have been working on AI for ten years”.
And rightly so: with 20 million job offers published on its site, 478 million visitors online and 6.6 million in its local agencies, the establishment seeks by all means to save its employees time. and advisors on numerous tasks, including classifying incoming images, automating CV analysis, processing emails and ensuring that job offers comply with legal standards. And this is what France Travail has chosen to focus on through its AI-oriented Employment Intelligence program, which includes around fifteen AI services in production.
Advertisement
An open source solution vs a paid solution
The organization has therefore developed a solution that meets this need called LegO. According to Agathe Ravilly, it brings several advantages, namely “replace a paid market solution, improve the quality of published job offers and reduce the time spent by advisors correcting job offers”. The platform developed on-premise (for obvious security reasons) is also open source and “thought for data scentists”assures Agathe Ravilly.
However, the teams in charge of the project encountered several challenges to arrive at this solution: in particular, it was necessary to achieve good model performance, effectively train large-scale models but also ensure the real performance of production and control the model in production. To respond to these challenges, the teams have therefore implemented a complete MLOps type life cycle. “in order to ensure scalable and secure deployment in production based on our AI and DevOps platform”, complete Agathe Ravilly.
Safety and performance at the top of the list
On the security side, Agathe Ravilly ensures that access to the data lake is secure and that the data is only handled within France Travail systems. “No data is transferred to a personal laptop. The platform also offers computing capabilities with CPU and GPU resources, resources which significantly reduce the time required to train complex models,” details Agathe Ravilly.
In this case, the team of data scientists opted for several LSTM (Long Short-Term Memory) models in order to achieve the requested performance. Result: training time went from 70 hours with CPUs to 20 minutes with GPUs. “For LegO, we have six LSTM models that address the 22 compliance criteria we defined. The first model, for example, addresses gender and age non-compliance”indicates Agathe Ravilly.
Advertisement
She adds that the data scientists have carried out significant preparation work. “They selected a comprehensive dataset that represents all non-compliance categories and scenarios likely to be encountered in real life. Approximately 50,000 offers were selected. The data was pre-processed until segmentation of offers into individual sentences, so they can be used to label offers and to train the model with more context.
First tangible results
The LegO service therefore entered production in 2021. “We process 600,000 new job offers per day in our system, and 20% of them are deemed non-compliant with the law by LegO,” specifies Agathe Ravilly, adding that the model achieved an accuracy rate equal to 82%.
She notes that the main non-compliance criteria identified are for example the request for good health, the request for a driving license or goods such as a car or even discrimination based on gender. “The service was available 100% of the time and its response time is around 300 milliseconds, thanks to our architecture. So we can say that it is a real success,” she concludes.
Obviously, several challenges remain to be overcome, in particular concerning the user interface is not adapted to this new AI service. Another sticking point cited, the return of use from advisors is not automatic. If they are faced with inconsistencies in the automatic processing of job offers, they still have to do it manually. “Today, the only way to control is to rely on our supervision data. We organize a manual validation phase every six months with a pool of advisors to check offline whether the prediction has always been correct or not “, indicates Agathe Ravilly.
“For example, a company may ask for a driver's license and a car, because there is no public transportation to the workplace. Asking for a driver's license and a car is considered discriminatory, “but knowing that there is no public transport is important. Advisors must therefore adapt job offers accordingly.”
Selected for you