France Travail relies on AI to validate the compliance of job offers

He was better known as Pôle Emploi. Renamed France Travail, the public administrative institution responsible for employment in France has not changed its mission. However, it seems that the organization is reaching a new level with the progressive adoption of new technologies, particularly 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 website, 478 million online visitors, and 6.6 million in its local agencies, the institution is seeking to save time for its employees and advisors in numerous tasks, including the classification of incoming images, the automation of CV analysis, email processing, and ensuring the compliance of job offers with legal standards. This is the focus of France Travail through its AI-driven Intelligence Emploi program which includes about fifteen production AI services.

An open source solution vs a paid solution

The organization has developed a solution to meet these needs called LegO. According to Agathe Ravilly, it brings several advantages, including “replacing a paid market solution, improving the quality of published job offers, and reducing the time spent by advisors correcting job offers. The on-premise developed platform (for obvious security reasons) is also open source and “designed for data scientists,” assures Agathe Ravilly.

However, the teams in charge of the project encountered several challenges to arrive at this solution: it was necessary, among other things, to achieve good model performance, effectively train models on a large scale, and also ensure the actual production performance and model control. To meet these challenges, the teams have implemented a complete MLOps lifecycle “to ensure evolutionary and secure deployment in production based on our AI and DevOps platform,” adds Agathe Ravilly.

Security and performance at the top of the list

In terms of security, Agathe Ravilly assures that access to the data lake is secure and that the data is manipulated only within the France Travail systems. “No data is transferred to a personal laptop. The platform also offers computing capabilities with CPU and GPU resources, which significantly reduce the time needed for training complex models,” details Agathe Ravilly.

In this case, the data science team opted for multiple LSTM (Long Short-Term Memory) models to achieve the required performance. As a result, the training time went from 70 hours with CPUs to 20 minutes with GPUs. “For LegO, we have six LSTM models that handle the 22 compliance criteria we have defined. The first model, for example, deals with non-compliance in terms of gender and age”, says Agathe Ravilly.

She adds that the data scientists did an important preparation work. “They selected a complete set of data that represents all categories of non-compliance and scenarios that may be encountered in real life. Approximately 50,000 offers were selected. The data was preprocessed up to the segmentation of offers into individual sentences, so that they could be used to label the offers and train the model with more context.

First tangible results

The LegO service entered production in 2021. “We process 600,000 new job offers per day in our system, and 20% of them are considered non-compliant with the law by LegO,” says Agathe Ravilly, adding that the model achieved a precision rate of 82%.

She notes that the main criteria for non-compliance identified are, for example, the request for good health, a driver’s license, or items such as a car, or discrimination based on gender. “The service has been available 100% of the time and its response time is about 300 milliseconds, thanks to our architecture. So we can say that it is a real success,” she concludes.

Of course, several challenges remain, including the fact that the user interface is not adapted to this new AI service. Another blocking point cited is the lack of automatic user feedback from the advisors. If they encounter 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, in deferred time, if the prediction has always been correct or not,” says Agathe Ravilly.

“For example, a company may ask for a driver’s license and a car, because there is no public transport to get to the workplace. Asking for a driver’s license and a car is considered discriminatory, but knowing that there is no public transportation is important. Therefore, advisors must adjust job offers accordingly”.

Selected for you


Leave a Reply

Your email address will not be published. Required fields are marked *