Government and Citizens
Artificial intelligence is an important technology for tackling the challenges in our society. We use AI to create a sustainable and prosperous digital future for our citizens. AI also enforces the digital governmental activities.
We demonstrate the AI research in following applications:
- In AI systems that are able to have a long-lasting personal dialog with citizens (dialog systems, conversational agents) and systems that can provide personnel recommendations (recommender systems). One application looks specifically into systems supporting recruitment.
- Citizens managing their personal data - more specifically, health data.
Special attention is paid to privacy, fairness, avoiding bias, and personalized and intuitive interaction.
Dialog Systems and Recommender Systems
Artificial intelligence for an Inclusive and Efficient Labour Market
Artificial Intelligence, and new technological evolutions sucjh as the Internet of Things, robotics, autonomous vehicles, biotechnology, 3D-printing,... all have a disruptive impact on the labour market. For many people the content of their job will change and this evolution is speeding up today. A study performed in 2018 estimates that in Flanders by 2030 4,5 million people will need significant training on the job, and another 310 000 employees will need complete re-schooling. Keeping the balance for an efficient job market in view of needs and skills of people will be very challenging. An extra diificulty is making the job market equally accessible to all work-seekers in view of cultural or demographic background.
The research within the Flanders AI Research Program
AI is partially causing the current shift in the job market, but it can also contribute to the solution of this challenge. Intelligent services can keep Flanders in a competitive position in the international context. We collaborate intensively with the public employment service of Flanders (VDAB) for different types of situations:
- Work-seekers: for example in providing better and fairer systems for job-recommendations and making these systems also available for advice on carreer-planning.
- Employers: for example by creating tools that allow employers to provide improved and more inclusive job-offerings.
- Policy makers: for example by providing tools for detecting unbalances in the job market, detection of any kind of discrimination or un-equal access to the job market and early detection of the need for new skills or educational programs.
More information
- Contact: Tijl De Bie, UGent/IDLab
- Research groups: KU Leuven (DTAI), UGent (DDCM, IDLAB, KERMIT), UAntwerpen (ADREM, AppliedDataMining)
- In collaboration with VDAB
Dialog systems - A Chatbot for Citizens on COVID Vaccination
In light of rapid evolving information streams, such as the current Covid-19-crisis, it is important to provide citizens with new and reliable information. A QA-bot is a chatbot that fulfils the need of people to ask their questions directly, without the requirement for them to search for the right answer themselves. The chatbot Vaccinchat can answer the most asked questions about vaccines and the vaccination strategy in Flanders.
Additionally, the evolution of questions throughout time provides important information to organizations about changes in the question for information itself. For instance, for vaccinchat.be this can seen at https://public.flourish.studio/visualisation/6517886/
The research within the Flanders AI Research Program
We investigate different manners in which natural language dialogs can play a role in innovative products and services. The QA-bots, like vaccinchat.be, are an example of this, but we also work on task-based dialogue systems to aid in problem solving or service execution (like a system that can book a hotel room or a system that can aid in explaining the steps in an industrial process).
There is a need for innovative research to cover different aspects of the problem: the understanding of natural language, the guidance of dialogues and the generation of the right answers. Within the research program, among others, we focus on reasoning, personalization, the use of background knowledge and multimodal systems that combine speech, images, and text. We make use of both classical and deep learning-based AI methods, as well as hybrid combinations. For instance, the vaccinchat.be QA-bot makes use of a novel deep learning “pipeline” that reduces the need of examples per question to train the system.
Results and Demonstrators
The chatbot www.vaccinchat.be (in Dutch) can answer the most frequently asked questions related to vaccination and the vaccination strategy in Flanders.
The evolution of questions over time provides important additional information, see e.g. tracking evolution of questions on vaccination.
Connect
- Contact: Walter Daelemans, (Uantwerpen, CliPS)
- Research groups on the chatbot for vaccination: UAntwerpen (CLiPS)
- Research groups on dialog systems: VUB (AI-LAB), UGent (IDLAB, LT3) en Leuven (LIIR).
More information
- Publications
· Maxime De Bruyn, Jeska Buhmann, Ehsan Lotfi, Walter Daelemans. BART for knowledge grounded conversations. Proceedings of KDD Workshop on Conversational Systems Towards Mainstream Adoption (KDD Converse’20). ACM, New York, NY, USA-2666 (2020) p. 1-6.
Dialog Systems - AI Explaining what it sees on a Picture
We investigate different manners in which natural language dialogs can play a role in innovative products and services. Dialogue systems can aid in problem solving or service execution (like a system that can book a hotel room or a system that can aid in explaining the steps in an industrial process).
One example of this is a (hybrid) visual dialoguesystem where a user can ask questions about what is shown on a picture.
The research within the Flanders AI Research Program
There is a need for innovative research to cover different aspects of the problem: the understanding of natural language, the guidance of dialogues and the generation of the right answers. Within the research program, among others, we focus on reasoning, personalization, the use of background knowledge and multimodal systems that combine speech, images, and text.
We make use of both classical and deep learning-based AI methods, as well as hybrid combinations. An example of the latter is the hybrid visual dialoguesystem where a user can ask questions about what is shown on an image. To be able to remember what has been said, the system stores information during the dialogue in its symbolic dialogue memory. The AI system is also capable of motivating why this answer was selected.
Results and Demonstrators
In the videos below, the memory of the AI system during a conversation is shown and you can see how the knowledge is continuously be completed. The AI system also explains how it can answer a certain question.
Demo 1: Visual Dialog
Demo 2: Visual Question Answering
Connect
- Contact: Katrien Beuls VUB, EHAI lab
- Research groups involved in this demonstrator: Evolutionary & Hybrid AI (vub.ac.be), (e-mail: ehai@ai.vub.ac.be)
- Research groups involved in the research on dialog systems: UAntwerpen (CLiPS), VUB (AI-LAB), UGent (IDLAB, LT3) en Leuven (LIIR).
More information
- Publications
· Nevens, J., Van Eecke, P., Beuls, K. (2019), Computational Construction Grammar for Visual Question Answering, Linguistics Vanguard, 5(1):1-16.
A Recommender System for Events
Every day, a large number of small and larger events take place for different types of audiences: youth, elderly, parties, nature-related, etc. Given this very broad offer of activities in Flanders, it can be sometimes quite hard to find interesting events. In practice, people often have no idea on certain activities close by and they thus miss out on events that would have liked to attend. Similarly, organizations miss the opportunity for additional revenues and publicity.
The goal of this demo is to build a recommender system for events, based on the user history and certain event parameters (location, targeted audience, price, type of events, …). We aim to achieve an bidirectional interactive recommendation system: at one hand, it should allow the user to put forward important event parameters that he finds important and should be considered when suggesting the next event. On the other hand, the system should explain its recommendations.
The research within the Flanders AI Research Program
When applying and researching AI, there is a growing need to have interpretable and explainable results. This system, additionally, also adds the interactive aspect to allow real-time user input about his/her preferences. This combination is novel and has not often been explored before. With the research and demonstrator we want to demonstrate that it is feasible to have a recommendation systems that complies to all the aspects mentioned above.
Connect
- Contact: Koen Ruymbeek en Bart Goethals (UAntwerpen)
- Research groups: UA-ADREM, UGent-Waves, UGent-IDLab, UGent-MICT, UHasselt-EDM, VUB-ETRO
- The data for the demonstrator is provided by publiq vzw
Citizens managing their personal data
Platform for the Exchange of Personal Health Data
We tackle the challenge of managing personal health data that is currently distributed over different places. We research how citizens can keep control over their data and can give permission to third parties to use their data.
The research within the Flanders AI Research Program
In the Flanders AI Resarch Program, the partners VITO, UHasselt, UGent andVUB collaborate on the research and development of a platform for management and sharing personal health data.
We use the ‘solid’ technology, in which citizens have their own data pod in which they can store their health data. We give the control over the peronal health data to the citizen.Citizens can manage their health related data preventively, before they become patients. Citizens can give permission to acadmeic partners, government and companies to use their data for research, which is beneficial for innovation and policy making.
The research includes the development of personal information agents who can exchange information with applications that have the required permission. This includes research on schema mappings: the integration of information from different sources in different formats and query rewriting:the automatic rewriting of queries sent by applications to the agents, depending on how the data is stored in the personal pods.
We also research in the set-up of the platform how to take into account the values of citizens and gain their trust in the data platform
Results and Demonstrators
The methods are applied to data collected by "de Gezondheidsgids" of Domus Medica, the association of general medical practitioners, in the frame of another project, BIBOPP (Burgers In Beweging met een Online Preventie Platform). BIBOPP is one of the applications that will make use of the data platform. For privacy reasons, the research program makes use of data of fictitious persons.
Connect
- Contact: Bart Buelens, VITO
- Research Groups: VITO, UHasselt (DBTI), UGent (IDLab), VUB (AI Lab)
Automated Document Analysis
Document Layout Analysis
Layout analysis is an important step in document processing. Examples include the analysis of invoices, pay slips, letters from suppliers and customers, or other business documents. With many companies still processing large volumes of scanned documents every day, there is great potential for AI engineers to automate this process and cut costs.
Today, OCR (optical character recognition) algorithms are usually used to convert scanned documents into text. These techniques can recognize individual letters (including handwritten letters) very effectively, but are unable to recognize the layout of complex documents, including figures, tables, graphs, columns, titles, etc. This creates many errors in the processing, and the manual work to correct these errors prevents many companies from adopting this on a large scale.
The research within the Flanders AI Research Program
The technology being developed within the Research Program is called “Document Segmentation with Probabilistic Homogeneity (DSPH)” and significantly improves the accuracy of existing OCR algorithms by first analyzing the layout of the documents. It uses a probabilistic method to do this, which hierarchically recognizes regions where text and other elements occur in a document, in the same way that a human parses the structure of a text. The algorithm takes into account geometric and other patterns.
Connect
- Contact: Ann Dooms (VUB)
- Research Groups: Digital Mathematics Research Group (VUB DIMA)
