Around the world, child helplines through their services provide a safe and confidential space for children to be heard and empowered. The Dutch Kindertelefoon is one of such helplines providing counselling services to children via call and chat all year round. In this thesis, we explore the design of a conversational agent for training counsellors of the Kindertelefoon. More specifically, we explore the design of an agent in a role-play setting where the agent acts as a child help seeker and the user, a counsellor of the helpline. We designed a conversational agent based on the Belief-Desire-Intention (BDI) model of agency that simulates a child victim of school bullying. Through interaction with the agent, a counsellor is able to learn the Five Phase Model, the conversation model that underpins the helpline's counselling methodology to ensure conversations remain child-centered. We tested a prototype based on this design with a group of counsellors at the Kindertelefoon with regards to their counselling self-efficacy and perceived usefulness of the system. Our results show that the conversational agent is able to influence the counselling self-efficacy of users, albeit a decrease in self-efficacy. The opposite would have been preferred for a learning tool to enable counsellors achieve more effective performance over time. However, feedback from participants indicate the potential of this conversational agent as an additional learning opportunity for training counsellors at the helpline.

In recent years, there has been an increasing number of patients with mental disorders. A conversational agent is being developed to ensure an easier diagnosis based on the chat between a patient and the agent. The objective of this research is to assess how well Support Vector Machine (SVM) classifies text into its corresponding schema, which are the mental states of the patient. In total, three different classifications have been attempted, Binary, Ordinal, and Per-Questionnaire. The experimental result indicated that SVM is possible to classify 2 out of 7 schema modes, but in general, the performance of SVM was not outperforming with a low f1-score. At the end of the research, SVM was compared to Recurrent Neural Network (RNN) and k-Nearest-Neighbour (kNN) and it turned out that RNN gives the best performance. One of the limitations affecting the result is the quality of the data set. With more correlated labels and a greater size of the data set, improved results can be expected.

People with diabetes often show symptoms heavily associated with depression, but these symptoms are often caused by the burden of daily diabetes management. The negative feelings caused by this burden are defined as "diabetes burnout''. Some of these negative feelings are caused by social issues. People with diabetes often resort to online sources to find out how to deal with these social issues since health care providers do not focus on this side of diabetes. Furthermore, the social stigma surrounding treatment for mental problems stops people with diabetes from going to a psychologist. In this project, a conversational agent is designed, implemented and evaluated to investigate whether it is capable of reducing social diabetes distress. The agent was designed to give personalized tips based on a social issue the person with diabetes shares. A longitudinal experiment was done over three sessions to evaluate the agent. The results show that the agent is able to reduce the diabetes distress more than a plain textual delivery of tips. The successful application shows the value of conversational agents and provide a basis to deploy such conversational agents in the e-mental health domain. The design we created can be used in future work, where a further personalized approach and a tool measuring the personalization should be investigated in order to better understand why the conversational agent is able to reduce the diabetes distress.