Challenge
On an annual basis, the Flemish Infoline ('1700') receives approximately 1 million telephone calls, 55,000 email messages and 25,000 chat calls. The process from question to answer is carried out 'systemically' with a number of linked components, including a (unique) editorial repository that feeds both the general website of the Flemish government (delivering content for 'Flanders.be') as the Flemish Infoline (delivering content and answer scripts for the operators of the Infoline), a dispatch system (CTI) that forwards the incoming calls to the ‘right' operator with the 'right' skills, and a customer tracking system (CRM) that registers all individual contacts (the question and answer scripts used).
The handling process of the Flemish Infoline (incoming question → answer script → outgoing answer) is meticulously recorded and reported. The answer scripts and the training of the operators are continuously adapted based on the reports, proceeding insights, internal feedback (from the operators on the scripts), etc.
However, each individual outgoing answer is still generated entirely 'manually' (recording the incoming question / searching for the correct answer script / editing and giving the answer by the operator). This means that the handling process of the Flemish Infoline is still quite long, while customers want answers to their questions ever faster and in a 'quasi-dialogue form'.
With this project Information Flanders wants to optimize the functioning of the Flemish Infoline with the help of Artificial Intelligence (AI) with an emphasis on Natural Language Processing (NLP), Machine Learning (ML) and Deep Learning (DL), and language and speech technology (Speech-to-Text [STT] and Text-to-Speech [TTS]). By using AI modules, the project aims to interpret and manipulate the question/answer process of the Flemish Infoline in order to make the whole process run more efficiently and automatically, using all the implicit knowledge present in the process. The project mainly aims at two objectives: (1) capture and categorize the question faster and more efficiently, and (2) forward the 'correctly captured' question faster and more efficiently to the 'right' operator with the 'right' skills, immediately sending a valid and standardized answer, so that the operator no longer has to look up the intended answer script.
It is not the intention to set up a separate AI project 'from scratch', but rather to integrate the AI enrichments into the existing (and already well-organised) process from question intake to answer output. NLP and ML can be conceived as autonomous modules/building blocks that act on the existing processes and enrich them with their own intelligence.
