SD Worx NLP
Belgian HR service provider SD Worx is active in 10 countries across Europe. Their payroll consultants daily spend 50 hours dispatching incoming client emails in Belgium alone. These emails need to be distributed across different departments each responsible for their own topics.
Belgian HR service provider SD Worx is active in 10 countries across Europe. Daily a lot of time is spent dispatching incoming client emails. These emails need to be distributed across different departments each responsible for their own topics.
A lot of manual work is involved in the distribution of those emails. To simplify that process and allow employees to free up time to perform other important tasks, we developed a mail classifier that uses artificial intelligence.
Thanks to Natural Language Processing we have a technology that is able to read emails and understand what the content is. In a nutshell, NLP makes it possible to train a model by feeding it texts until it understands what’s written.
We built a state-of-the art language model for SD Worx, based on the destination, body and subject of emails. This model returns a prediction for each email that gets fed into it. This prediction is then used to forward the email to its respective destination.
SD Worx receives emails in both Dutch and French, so we had to use a model architecture that was capable of handling these two languages. If you want to know more about the multilingual side of this project, feel free to read our technical blogpost about the topic.
The result is a REST API that SD Worx can plug into its general inboxes. At the moment our algorithm is able to forward ¾ of the mails automatically, leaving a lot more time to focus on the real thing: helping customers. In the future we want to implement a feedback loop in this process that allows the algorithm to learn from its mistakes, so we can potentially improve its performance along the way.
Based on our model, we’re able to forward incoming mails to the right service or group of people, in the near future we want to be able to predict to which person the mail belongs.