This paper presents a deception-detection module for a diabetes support system, addressing the challenge of unreliable patient self-reporting and ultimately attempting to improve diabetes care. The research is for a system called CHIP developed by the Hybrid Intelligence project group and TNO. Linguistic cues, such as motion verbs, negation terms, and exclusive terms were identified through a literature study and encoded using custom dictionaries. Cue detection was implemented using the SpaCy NLP library, which identifies and counts cue occurrences. A stylometric machine learning approach was favored over LLMs for explainability and scientific substantiation. In this research, an SVM model, selected for its alignment with prior research (the Mafiascum experiment), was trained on annotated Mafia game data, using normalized cue frequencies as features for the model. Although the SVM achieved high accuracy on truthful messages (F1 between 0.78–0.84), it performed poorly in detecting deception (F1 between 0.21–0.22), likely because of the high frequency of truthful input compared to deceptive input. The low accuracy, along with the model’s domain transferability and performance limitations, suggest further work is needed, particularly with context-specific data and possible integration with LLM-based approaches.

Lifestyle management systems aim to provide personalized health guidance by interpreting patient's self-reported data. However, these systems often overlook the temporal consistency of behavioral patterns, risking inaccurate or misleading recommendations. To address this, we present an adapted version of a diabetes lifestyle management system that integrates a conflict detection component into its knowledge graph (KG) architecture. This addition enables the system to identify temporal inconsistencies in patient data based on formally defined rules. We developed handcrafted constraints to capture clinically meaningful relationships, such as the expected timing between physical activity and glucose responses. A tunable threshold parameter, $\varepsilon$, is introduced to account for minor fluctuations and variability in activity intensity, enhancing the physiological realism of constraint evaluations. Handcrafted rules were favored over mined constraints due to the limited availability of real-world data, though future work may incorporate mining techniques when larger datasets become accessible. The system was tested on synthetic case studies that simulate typical lifestyle scenarios, demonstrating its potential to flag inconsistencies and improve the reliability of temporal reasoning. These results contribute to the design of more trustworthy, context-aware decision support tools for personalized health management.

Deceptive self-reporting in diabetes lifestyle management (DLM) systems limits their ability to offer meaningful and accurate support. Deception can function as a self-protective mechanism, driven by factors such as low self-esteem or the desire to protect self-image. This research builds on CHIP, a chatbot-based DLM prototype, to explore whether the language framing of its responses can influence the psychological determinants of deception. Two framing strategies, empathic and affirming, were implemented and evaluated through a pilot user study, which offers insights for refining the intervention and experimental design in future research.

Under which conditions can you say that a system is actually and meaningfully under your control? Accidents happen with machines and often that is not the fault of the user. So what does control entail? To gain some modicum of understanding, we need to learn how technology and control relate to one another. Of course, the concept of control also relates to the consequences of control, namely the responsibility you have for that which you can control.
On average,we are used to thinking about technology as a kind of hammer, something on hand that we can make use of. Yet, such a hammer also invites us to hammer things. This invitation is a kind of mediation through which we are encouraged to act in one way and not another. In short, technology can also influence us.
If technology can exert a kind of influence on you, then we need to ask what that means for control. What kind of issues do we run into because of that influence? In this dissertation I investigate a few key issues: self reinforcement and relevancy. Technology tends to entrench itself in society once it is widely implemented. This process of entrenchment is often through self-reinforcement (chapter 2). Like a snowball rolling down a hill that picks up more snow as it goes, so too can technology gain a kind of traction that becomes harder and harder to ignore and disband or even change. Consider, for example, how much has to change if we want to live without cars. The moment technology gets picked up at large, we also institute policies and create institutions around which such technology can be legitimized. The point is that the technology can create a new standard to which everyone grows accustomed...

This thesis investigates the role of learned abstract models in online planning and model-based reinforcement learning (MBRL). We explore how abstract models can accelerate search in online planning and evaluate their effectiveness in supporting policy evaluation and improvement in MBRL.

In online planning, we focus on reducing the high computational cost of simulating large, factored, partially observable environments. In Chapter 3, we introduce the influence-augmented local simulator (IALS), which approximates external influences while preserving local agent interactions. By replacing the full simulator with IALS, we enable faster planning while maintaining decision quality. We propose a two-phase approach where the influence model is trained offline and later integrated into planning, allowing significantly more simulations within a fixed computational budget. However, this approach has limitations, including potential distribution shifts and the risk of poor generalization.

To address these issues, Chapter 4 introduces the self-improving simulator, which eliminates offline training by learning the abstract model online during planning. A simulator selection mechanism dynamically balances the use of the learned and original simulators, improving computational efficiency over time while ensuring planning accuracy. Our results show that this approach avoids distribution shift issues, prevents premature reliance on inaccurate models, and removes the delay associated with offline training.

In MBRL, we examine the effectiveness of MuZero’s learned model in supporting policy evaluation and improvement. In Chapter 5, we analyze how well MuZero’s model generalizes beyond its training distribution and find that it struggles to support planning "outside the box" due to accumulated model inaccuracies. However, we show that MuZero’s learned policy prior mitigates these errors by guiding the search toward regions where the model is more reliable. This insight highlights the dual role of the policy prior—not only improving search efficiency but also compensating for model imperfections, contributing to MuZero’s strong empirical performance.

Overall, this thesis advances the understanding of learned abstract models in sequential decision-making, demonstrating their potential to improve computational efficiency while identifying key limitations in their ability to support planning. We hope these findings encourage further research into abstraction-driven approaches for adaptive, scalable decision-making in complex environments.

The effects of digital technologies on freedom and democracy have garnered increasing attention in recent years. Many have raised concerns about surveillance capitalism, technofeudalism, and general threats to constitutional democracies—with a special convergence on the worry that uncontrolled power of online platforms undermines people’s freedom. However, it remains unclear how ‘freedom’ should be understood, what the relation is between freedom and uncontrolled power, and to what extent these worries extend beyond online platforms. In this dissertation, I argue that these problems are best answered by appealing to a neo-republican account of freedom as non-domination, where ‘domination’ is understood as a condition of living under an agent’s uncontrolled power. In the context of AI systems used in core societal sectors such as healthcare, I show that domination of a system’s (in)direct end-users by the system’s developers occurs in at least three ways: (1) the distribution of decision-making power, (2) technical limitations of AI systems, and (3) underlying societal structures that empower developers and disempower end-users. To safeguard freedom in the digital age, I propose that AI development requires the explicit intention to 'design for non-domination'. This requires us to consider the broader societal contexts within which these systems operate, such as current regulatory initiatives and the political economy.

This research revolves around measuring the quality of arguments. High-quality arguments help in improving political discussions, resulting in better decision-making. Wachsmuth et al. developed a taxonomy breaking down argument quality into several dimensions. This work makes use of that taxonomy and combines it with modern NLP models. A cross-dataset examination of argument quality models was conducted. In particular, models were investigated on their generalizability between dimensions. Overall results show that there is no large difference in accuracy and agreement when models predict data of a quality dimension they were trained on, over dimensions they were not trained on. One can conclude that generalizations of argument quality dimensions with language models were not found. Nevertheless, qualitative analysis highlights findings that indicate some generalization to other dimensions.

Moral values influence humans in decision-making. Pluralist moral philosophers argue that human morality can be represented by a finite number of moral values, respecting the differences in moral views. Recent advancements in NLP show that language models retain a discernible level of knowledge in deontological ethics and moral norms of society. However, a model which can only decide either right or wrong cannot fully understand the diverse moral perspectives of humans.

We propose a moral sentence embedding space, which can encompass moral differences, through the state-of-the-art Contrastive Learning framework. We evaluate the moral embedding space both intrinsically and extrinsically via three tasks: classification, moral similarity, and visual analysis. We show that our moral embedding space understands the characteristics of each moral value. Our results also highlight that moral rhetoric is seldom explicit in the text, emphasizing the necessity of additional information such as moral labels.

Large Language Models (LLM) have brought significant performance increase on many Natural Language Processing tasks. However LLMs have not been tested for meeting summarization. This research paper examines the effectiveness of the gpt-3.5-turbo model on the meeting summarization domain. However due to input length limitations, it cannot be applied directly to this task. Thus the paper investigates two segmentation methods: a simple context-length window approach and topic segmentation using Latent Dirichlet Allocation (LDA). The context-length window approach's performance is close to the Pointer Generator framework. The topic segmentation gives worse results. Overall gpt-3.5-turbo performs worse with both approaches than state-of-the-art models which use a transformer architecture adapted for long documents.

This paper presents a novel approach to detect agreement and disagreement moments between participants in meeting transcripts without relying on labeled data. We propose a model in which disagreement detection is defined as the process of first identifying argumentative theses relevant to a given corpus of text and then classifying all phrases in the text as being either in favor of, against or expressing no opinion on a given thesis. To identify relevant theses, we compare the performance of a latent Dirichlet allocation-based topic model against that of a diverse set of large language models. To classify the stance of a phrase with respect to a thesis, only large language models are used. We find that, while state-of-the-art large language models do not outperform topic modeling-based approaches in extracting semantically relevant content, they are capable of presenting such content in a more concise and grammatically correct manner. We also find that state-of-the-art large language models are not capable of accurately performing stance classification as described above.

Meetings are the keystone of a good company. They allow for quick decision making, multiple-perspective problem solving and effective communication. However, most employees and managers have a negative view on the efficiency and quality of their meetings. High quality meetings where every participant feels equally heard and respected is crucial for having positive meeting sentiment within a company. One of the most influential aspects of meetings are speech overlaps. Overlaps range from short utterances such as backchannels, to follow up questions and clarifications, to complete interruptions. In non-competitive cases, the overlapped speaker feels that the other participants are listening and actively engaging with them during the meeting. In competitive cases, the overlapped speaker can feel interrupted and unimportant. Therefore, competitive overlaps often have a negative impact on the course of the discussion and the overlappee's meeting sentiment. In problematic cases, these overlaps should be reduced to a minimum. In order to do this, overlaps must be classified as either competitive or non-competitive. This paper proposes a novel approach to overlap classification, namely that of text-based classification through Large Language Models. Four different prompt designs are used and tested on the two best performing and publicly available models, GPT-3.5-turbo and GPT-4. The results show that the in-context learning approach using the GPT-4 model results in the most accurate classifications. When comparing the results to previous work, it is observed that the text-based GPT-4 model matches carefully engineered neural networks that even adopt a multi-modular approach.

In April 2020, a Dutch research team swiftly analyzed public opinions on COVID-19 lockdown relaxations. However, due to time constraints, only a small amount of opinion data could be processed. With the surge of popularity in the field of Natural Language Processing (NLP) and the arrival of tools like ChatGPT, a number of tasks involving Large Language Models (LLMs) have become easier. This study aims to address the effectiveness of these LLMs on stance detection using this COVID-19 opinion corpus. The corpus is chunked and sampled to be used as input for OpenAI's GPT-3.5-Turbo LLM. The machine-generated stances are then evaluated against multiple binary classification metrics. It is shown that these models perform very well in the field of stance detection, with an average F-score of 0.895. However, a significant number of misclassifications are observed in one dataset. Therefore we conclude that while LLMs offer valuable guidelines, it is still crucial to verify their outputs when dealing with complex or important public matters.

Behavior support applications require an understanding of the user's values to provide personalized support. User models have been developed to capture this relationship, but they may not always address unforeseen circumstances due to the changing nature of human values. To address this, the ability to update value models in real-time is crucial. This paper presents a study on the design of a graphical user interface (GUI) that utilizes behavior trees to elicit and update personal values. The study focuses on comparing the accuracy and usability of different questioning techniques, specifically comparative questions and isolation questions, and user interfaces. The results show that GUI with isolation questions and text user interface with comparative questions performed better in terms of accuracy compared to GUI with comparative questions. The study also highlights the possible impact of discrepancies in experimental procedures on the results. Furthermore, the limitations of the behavior tree model and the need for model improvements are discussed. Future work is proposed to address these limitations and explore enhancements such as alternative models, supervised processes for baseline tree building, frictionless elicitation techniques, consideration of random value shifts and big life events, and the potential of large language models. Overall, the findings contribute to the understanding of eliciting and updating user value models in personalized systems.

Since the recent rise and advancement of video conferencing platforms such as Zoom, it has become important to interpret the logistics of remote online meetings. Analysing verbal and non-verbal cues (such as body language) between members of these virtual forums can provide additional information regarding the level of conversational involvement of each party. This research focuses on age, gender, demographics and virtual background differences in the context of group conversational discussions. It argues that groups formed of younger adults have a higher level of involvement compared to the older groups. Similarly, this study found that groups with a higher ratio of male participants score better in virtual conversational engagement compared to women preponderant groups. To better understand the influence of these inter- and intra-personal characteristics, a corpus formed of 45 online meetings on the topic of Covid-19 was used. This set of data consists of questionnaires with measurements (demographics and other personal values), as well as detailed annotations of conversational signals, which provide valuable insights into the research topic of conversational involvement.

This study includes an experiment to investigate the involvement of individual backgrounds in the prediction of group conversational engagement. Four predictive models were used, namely the Decision Tree, Random Forest, Linear Regression and Generalized Linear Mixed Effects Models. While the Generalized Linear Mixed Effects Model provides more meaningful observations on the statistical effect of these factors, the Random Forest ultimately proved to give the best performance accuracy. The purpose of this research is to improve the connection between humans and technology by studying how inter- and intra- characteristics of individuals impact the involvement of a group in virtual interaction.

Involvement in an online setting has become increasingly more important in a rapidly digitizing world. Previous research has suggested that lexical alignment is an effective method for increasing interpersonal involvement. Multi-modal systems exist for analysing involvement, but the lexical alignment modality is not commonly used in these. This study examines the lexical alignment modality as a predictive measure of involvement in online group meetings. By analyzing participation indicators in virtual meetings, the aim is to identify moments of high involvement. Results indicate that lexical alignment does not correlate with moments of high involvement. The findings contribute to the scientific understanding of leveraging the verbal modality for analyzing and organizing online meetings.

The importance of understanding how to maximize involvement in virtual meetings has been greatly increased due to the rapid rise of video conferencing tools during the COVID-19 pandemic. This research builds upon the data collected by the MEMO Corpus, specifically interview footage and annotated data, and will aim to explore the various factors that influence group involvement in virtual meetings. The research will be conducted in three steps; data annotation, data exploration, and data analysis. The goal of this paper is to investigate the influence of factors such as likeability, listening, and audience engagement in how conversational ability is perceived, given that the scales provided by the MEMOCorpus are valid and the three measures are independent, and how conversational ability influences group involvement, by exploring the existence of correlations. The results of the analysis will provide insight into the dynamics of conversational involvement in virtual meetings and improve virtual communication practices.