Offensive-language detection is often evaluated as a single-label classification task, although research on subjective annotation shows that people can make systematically different judgements about the same text. This thesis focuses on predicting annotator-specific offensiveness judgements rather than establishing an objective offensiveness label. Moral values are examined as one possible personalization signal because they provide an individual-level description of normative concerns that may be relevant when people interpret potential violations of social and moral norms. The study asks whether different representations of the same measured moral-value profile change an LLM’s approximation of an annotator’s observed binary judgement.

The experiments use D3CODE, which combines repeated individual offensiveness annotations with annotator-level scores on six moral foundations. Five deterministic renderings of each profile—natural language, categorical, numeric, ordinal, and top-k—are compared with a no-profile, unpersonalized reference. Qwen2.5-72B-Instruct is used as the primary model; Qwen2.5-7B-Instruct and Llama-3.3-70B-Instruct provide within-family and cross-family robustness comparisons. The models are evaluated in zero-shot prompting and in three same-annotator few-shot settings with k = 3 examples selected randomly, by semantic retrieval, or by diversity. Evaluation combines Accuracy, Macro-F1, paired McNemar comparisons, and an instance-level cross-representation decision-disagreement diagnostic.

The zero-shot condition exhibits the largest spread in aggregate performance across profile representations and cross-representation disagreement above 20% for each selected model. Same-annotator few-shot context is associated with lower disagreement, although profile representations continue to change individual predictions. Retrieval-based prompting produces the strongest Macro-F1 across the selected models and the lowest disagreement in most sub-categories for the primary model. No representation performs consistently best across every tested model, setting, and metric. Instead, the results show that profile representation and demonstration selection are consequential parts of the experimental treatment and should be reported, controlled, and evaluated in personalized, value-conditioned judgement prediction.
... Read more

“Don’t tune hyperparameters on the test set” is often stated in machine learning textbooks. Violating it is considered a cardinal sin that produces misleadingly optimistic results, corrupts benchmark integrity, and thus can even be interpreted as scientific fraud. Yet evidence suggests that test set hyperparameter tuning does occur in practice, making it all the more important to understand its actual consequences. So how bad is it, really? In this work we question this dogma and put it to an empirical test. We systematically study the magnitude of the performance inflation caused by tuning the hyperparameters on the test set for MNIST-1D and CIFAR-10. Our experiments show that while the effect is real and significant, it is frequently small relative to other sources of noise. In many cases, we find that tuning on the test set recovers exactly the same model as when tuning on the validation set. Most importantly, we find that the rankings of models remain preserved after tuning on the test set, and therefore that consistent test-set tuning does not invalidate benchmarks or model selection. Our results call for a more nuanced view of tuning hyperparameters on the test set, stimulating researchers to openly report test tuning. ... Read more

Automatic speech recognition (ASR) systems achieve very low word error rates on standard adult speech but underperform on populations whose speech deviates acoustically from the training distribution. Dysarthric speech, in which neurological injury impairs the muscular control of speech production, is among the most challenging cases: word error rates can exceed 90% on severely impaired speakers even after in-domain fine-tuning. Mixture-of-Experts (MoE) architectures, in which feed-forward sublayers are replaced by parallel “expert” sublayers and a small gating network routes each input to a subset of them, offer a candidate mechanism for inclusive ASR: in principle, the experts could come to specialize on different speaker populations and the router could allocate them where they are needed.

This thesis investigates whether MoE delivers that promise on dysarthric ASR, using the FastConformer-CTC Large encoder fine-tuned per fold on the TORGO and UASpeech corpora across 41 leave-one-speaker-out folds. Three research questions are addressed: how standard sparsely-gated MoE compares to a fine-tuned dense baseline (RQ1); what specialization the trained experts develop (RQ2); and whether either training-time supervision or inference-time router steering can push the model toward useful specialization (RQ3).

The answer to all three questions is negative. Five standard MoE configurations match but do not beat the fine-tuned dense baseline, and a parameter-equal follow-up shows that at equal feed-forward capacity, the MoE structure is actively worse than dense by up to 2.89 pp. The trained experts develop genuine functional specialization, but the axis of specialization is encoder depth, not speaker group or phonetic content: different layers prefer different expert pairs, and within any layer, routing is roughly content-independent and group-independent. Training-time supervision toward group-aware specialization collapses at inference, as the depth-based organization overrides the supervised partition. Inference-time router steering produces real per-speaker improvements under post-hoc selection of the bias strength, but neither of two deployment-realistic mechanisms for choosing that bias recovers the gain.

The MoE structure on this task does meaningful work, but not the work the task needs: it organizes the encoder’s computation along an axis the dense baseline already covers through its per-layer feed-forward weights. Routing-input variants, cross-architecture replication, and continual-learning extensions are flagged as the most direct follow-ups. ... Read more

This thesis examines whether Dutch adult dysarthric speech can be transformed into speech that is sufficiently child-like to improve Automatic Speech Recognition (ASR) for Dutch children with Developmental Language Disorder (DLD). Due to the limited availability of annotated speech data from children with DLD, existing ASR systems often struggle to accurately recognize this population. To address this data scarcity issue, this work explores the use of traditional Digital Signal Processing (DSP) techniques to augment recordings of Dutch adult speakers with dysarthria to sound more child-like, while preserving overlapping characteristics with DLD.

A series of transformation methods targeting pitch, vocal tract characteristics, spectral properties, and prosodic patterns are implemented and combined into multiple transformation pipelines. The resulting synthetic speech is evaluated through ASR fine-tuning experiments using Whisper-large-v3 and Word Error Rate (WER), as well as an objective speech-quality assessment using Automatic Mean Opinion Score (AMOS) prediction.

The results show that the synthetic child-like speech can indeed improve ASR performance when used as augmentation data. The best-performing configuration combines a targeted subset of mild and moderate hypokinetic and ataxic dysarthric speakers with a prosody-enhanced transformation pipeline, and achieved a WER of 29.00\%, outperforming both the zero-shot Whisper-large-v3 baseline (46.48\%) and a CHILDES-only fine-tuning baseline (31.57\%).

The findings further indicate that source-speaker similarity is more important than dataset size and that conservative prosodic modifications contribute meaningfully to the usefulness of synthetic speech data for ASR fine-tuning. Overall, the results demonstrate that carefully selected dysarthric speech combined with interpretable DSP transformations can provide valuable augmentation data for Dutch children with DLD. ... Read more

Automatic speech recognition (ASR) is increasingly used in speech-based services,
but its performance is often worse for children and non-native speakers. This study
examines whether utterance length helps explain the higher word error rates observed for non-native Dutch-speaking children compared with native Dutch-speaking children, and whether this relationship differs between read speech and human-machine interaction (HMI) speech. Using 42,415 utterances from 188 child speakers in the Dutch Jasmin corpus and ASR outputs from Google Chirp 2 and Whisper-large-v3, the analysis compares word error rates by utterance length, measures native/non-native performance gaps, standardizes both speaker groups to the same length distribution, and estimates uncertainty with speaker-level bootstrap resampling. The results show that shorter utterances have higher error rates for both groups, speech types, and systems. Utterance length explains a measurable part of the read-speech disparity, because the gap decreases with length and is reduced after length standardization. However, a substantial gap remains. For HMI speech, the evidence is less consistent. Overall, utterance length affects how ASR disparities are measured, but it does not fully explain the disadvantage for non-native child speech. ... Read more

Recent developments in Artificial Intelligence offer new possibilities for the development of systems which adapt to human motivation or emotion. These can have a variety of applications such as making therapy more accessible or boosting student motivation or engagement. In order to gain an overview these applications, of adaption strategies and inputs are used and challenges researchers are facing during development, a Systematic Literature Review was conducted. The review follows PRISMA guidelines and includes from Scopus, Web of Science and IEEE Explore. Due to time limitations, only papers published in 2024 and 2025 were considered. The final review includes 139 papers, out of which only 6 target motivation. 46 were chatbots, and 53 were recommendation systems. Researchers primarily struggled with the interpretation of complex emotions, and the need for more adaptation options. Lastly, a strong need for extensive user testing and comprehensive data privacy measures was identified. ... Read more

This paper is a systematic literature review (SLR) investigating how intelligent systems leverage learning-related information to adapt their behaviour. This paper is done according to PRISMA guidelines, which ensures reproducibility. For this review, we analysed 58 papers published after 2023. The review focused on types of inputs, system objectives, and application domains. The systems in this survey adapt their behaviour based on various inputs, such as facial expression, gestures, eye gaze, and user preferences. These systems enhance the learning experience by increasing user engagement and motivation. Many of these papers target the education domain, such as STEM, but there are also papers focused on motor skills training or cognitive training. Time constraints limited the scope of the review, particularly in identifying long-term trends. But the result can be a solid base for future research into adaptive learning and training platforms. ... Read more

Congressional hearings are at the center of legislation, yet their analysis is hindered by the volume and complexity of the transcripts. While recent advances in Natural Language Processing (NLP) have enabled political discourse analysis using automated tools, conventional topic modeling methods often struggle to produce semantically coherent topics due to their reliance on context-free word frequencies. This paper evaluates the performance of a new transformer-based topic modeling technique, focusing on its application to policy discussions through a detailed case study. Two variants of BERTopic are considered: (1) a parameter-tuned model and (2) a zero-shot variant, evaluated on U.S. congressional hearing transcripts from 2021 to 2024. The results demonstrate that the zero-shot version achieves competitive coherence with increased interpretability and stability, making it a useful resource for policymakers and researchers alike. This paper establishes a foundational methodological framework for automated legislative text analysis. It also outlines the trade-offs between unsupervised and semi-supervised topic modeling in political usage. ... Read more

A well-functioning democracy depends on an informed population. To help informing citizens, summaries of arguments in political transcripts can be made. An approach to argument summarization is the creation of summaries through distillation of the arguments into higher-level key points. In this approach, mapping arguments to key points is an important subtask. This study examines how model selection, prompting strategy, choice of domain, and input batching influence the performance of large language models (LLMs) in matching arguments to key points. We introduce a self-annotated dataset from U.S. Congress committee transcripts and evaluate both generative and embedding-based models on this task. Generative LLMs (GPT-3.5-turbo, o4-mini) outperform both untuned and fine-tuned RoBERTa in zero-shot argument-to-keypoint mapping (up to 0.880 macro-F1), while sparse two-shot prompting yields no gains. Moderate batching (n=32) boosts throughput without losing accuracy. These results show that a fully automated KPA pipeline—argument extraction, key-point generation, and mapping—is achievable with current LLMs. ... Read more

With an increased demand for personalized systems, adaptive systems can support its users by recognizing their cognitive state, and adapt different elements to improve the user's mental state. With a literature survey, the following question has been answered: how do intelligent systems acquire and use information related to user attention? The answer was found by following the PRISMA guidelines, using the identification flowchart, to ensure reproducibility of the identification and screening of the papers, in addition to applying consistent criteria to which articles could be included in the review. This process resulted in 74 papers that fit the criteria. The results showed a large variety in the sensor input, modeling, objectives and domain, while the adaptation strategies could be summarized by five categories: UI change, feedback timing, automation level, difficulty adjustment and behavioral feedback. Combinations of categories were also present. UI changes and feedback timing were the most popular categories, especially from 2015 onward. Difficulty adjustments were surprisingly rarely utilized, especially in articles focused on education, possibly because the adaptation requires additional complexity to be added to the system. Challenges described in the literature were mainly focused on short term improvements, instead of long term issues. ... Read more

Visual impairment affects over 2.2 billion individuals globally, emphasizing the critical need for effective assistive technologies. This work focuses on developing a video captioning model explicitly tailored for visually impaired users, leveraging advancements in deep learning techniques. Video captioning involves converting video frames into textual descriptions, effectively bridging the domains of computer vision (CV) and natural language processing (NLP). We surveyed young visually impaired individuals from the Visio organization, who provided key insights into the design of our model.
We enhance the existing S2VT model by modifying its temporal attention mechanism to improve the recognition of visual surroundings, addressing the unique challenges visually impaired individuals face.
This research explores critical questions surrounding the model's sensitivity to actions, the readability of generated captions, and methods for latency reduction. To evaluate the model's effectiveness, we implement readability metrics—an approach not previously utilized in video captioning assessments. Our findings contribute to enhancing accessibility and independence for visually impaired individuals through advanced video captioning solutions. ... Read more

Training deep learning models for time-series prediction of a target population often requires a substantial amount of training data, which may not be readily available. This work addresses the challenge of leveraging multiple related sources of time series data in the same feature space to improve the prediction performance of a deep learning model for a target population. Specifically, we focus on a scenario where the target dataset, representing the desired target population, is underrepresented, while the source datasets consist of mismatched populations that are sufficiently representative for training a deep learning model. In this study, we explore state-of-the-art techniques, including transfer learning, ensemble learning, and domain adaptation to leverage source datasets towards a target population using real-world medical data. Additionally, we investigate the use of model performance-derived baselines as a heuristic to quantify the magnitude of the distribution mismatch between a source(s) and a target. Our results demonstrate that a set of well-defined baselines can effectively quantify the distribution mismatch and provide insights into the choice of leveraging technique for a given mismatch scenario. Furthermore, our results show that all state-of-the-art techniques can be employed to leverage related source datasets towards the target, though the performance of these techniques varies depending on the characteristics of the distribution mismatch. Eventually, we discuss the applicability of this research to new scenarios, along with avenues for future research. ... Read more

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. ... Read more

Artificial social agents (ASAs) are systems designed to interact with humans in a socially intelligent manner. As the field of robotics is rapidly advancing, some studies focused on creating more effective agents by analysing how people perceive them. However, culture affects people’s perception of ASAs. So, cultural aspects must be examined in order to create more effective ASAs. This study aims to contribute to the research of cultural influence on ASAs by discovering the cultural similarities and differences in the perception of ASAs from Dutch and Chinese speakers. An English questionnaire has been made to evaluate ASAs and also has been translated into Chinese. In this study, the questionnaire is translated into Dutch and validated using both the original and the translated questionnaire. While assessing the translation, data is gathered from Dutch speakers on a human-ASA interaction. This data is then compared to the previously collected data from the Chinese translation study by calculating their mean, standard deviation and t-distribution of the differences between both cultural groups. On an item level, the results show a satisfactory level of correlation (ICC M = 0.67, SD = 0.12, range [0.36, 0.92]). On a construct/dimension level, results show a good level of correlation (ICC M = 0.80, SD = 0.11, range [0.52, 0.93]). Correction values for the translation have been advised for converting item scores between the two questionnaires. Cultural differences have been found and reported between Dutch and Chinese speakers, which could be used in future research on creating more effective ASAs. ... Read more

Humans interact with various Artificial Social Agents (ASAs) on a daily basis. ASAs range from the Honda robot ASIMO to Apple’s Siri. To measure the perception of human-ASA interactions, a standardized questionnaire was created. Yet, this questionnaire was so far only available in English and Chinese. It has been found that culture can affect how these interactions are perceived. The aim of this study is to answer the question: What are the differences and similarities of the English and German human-ASA interaction interpretations? In this paper, we translate the questionnaire into German, validate it. Once proven valid, we give the English and German questionnaire on bilingual participants who watch a human-ASA interaction video and rate it in both languages. We measure the differences and similarities between the English and German responses. At the end, we combine the finding from the questionnaire results with examples from literature to form recommendations for future ASA developments. We conclude that an average good level of correlation between the two languages for the 90 questionnaire items (ICC M = 0.65, SD = 0.14, range [0.27, 0.90]), on the construct level (ICC M = 0.8, SD = 0.1, range [0.51, 0.92]), and for the 24 representative items (M = 0.67, SD = 0.14, range[0.31, 0.90]). Additionally, we found systematic differences between the English questionnaire scores of the bilingual sample seen in this study and a previously established mixed-English sample. ... Read more

The Artificial Social Agent(ASA) questionnaire serves as a tool to assess the interaction between humans and ASA\cite{OGQuestionnaire}. In an effort to increase the number of people who can make use of the questionnaire, previous work has been done to translate this questionnaire to Mandarin Chinese\cite{Chinese}. To increase the number of people who can use this questionnaire even more, we have translated it to German using a similar iterative translating process. We had professionals translate the questionnaire and ran 3 rounds of surveys to see what questions needed improvement. Then, there was a final survey that evaluated the German version to see how well it performs and to see how it compares to the English version. In order to evaluate the cultural differences between the Chinese and German translations, the data of both translation evaluations were compared. From these results, it was shown that ASA developers who use the German version of the ASA questionnaire can make comparisons on a construct level to the English version. This was, however, only possible when using the full questionnaire, not the short version. Additionally, we have found that there are differences in how Mandarin Chinese speakers and German speakers rate ASAs.
... Read more

The work presented in this thesis investigates the creation of virtual sound sources in a room equipped with a limited number of loudspeakers. This limited number of loudspeakers is typical for consumer loudspeaker systems. Ideally, these systems can provide a listening experience in which localisation cues are completely present. However, in current systems, this is not the case. The limited number of loudspeakers make it impractical to produce a physically accurate sound field. A possible solution is to create a perceptually accurate sound field instead. In this work, a step towards an algorithm which can do so is presented. The developed algorithm requires knowledge of the listener placement, the room, and the loudspeaker placement. Spatial weighting is used to construct a region in which the acoustic en-ergy should be large and a region in which the acoustic energy should be limited. The loudspeaker playback signals are obtained by maximising the energy ratio between these regions, while at the same time ensuring that the perceptual difference between the received audio and reference audio remains limited. The algorithm employs a convex optimisation problem to facilitate efficiently solving for the playback signals. Concretely, six convex optimisation problems are proposed with somewhat increasing complexity and different weighting matrices. For each of the optimisation problems, the proposed algorithm is compared against a simple amplitude panning algorithm and a nearest neighbour algorithm. It is found that none of the considered algorithms is clearly preferred over the others in terms of the considered evaluation metrics. A major limitation of the presented work is that the evaluation metrics do not explicitly test for localisation accuracy. In future work, this should be investigated by including subjective tests. ... Read more

Over the past decade, spatial audio awareness evolved into an in-demand feature in audio entertainment. The addition of sound source locations to, for instance, movies or music adds a level of auditory envelopment and spatial awareness to the audio experience. Expensive setups present in, for instance, cinema's, are able to create this envelopment by means of a large set of loudspeakers with which the desired sound fields are created. Creating this spatial envelopment in practical consumer living rooms or home cinema setups proves to be a more challenging task due to the impractical amount of loudspeakers required. To create the audio envelopment with a small amount of loudspeakers, crosstalk cancellation can be used. Crosstalk cancellation as posed in literature is, however, not robust enough to be used in practical appliances. The main cause of these bad characteristics is the objective cost function it optimizes which results in an ill-posed problem. In this thesis, the crosstalk cancellation problem is relaxed by aiming for perceptually sufficient results instead of aiming for objectively optimal results. The human auditory system has its limitations in both the perception of audio and localization of audio sources. Exploiting the limitations of the auditory system generates mathematical freedom that can be used to construct a more robust and stable crosstalk cancellation algorithm. This thesis provides the first steps towards the incorporation of audio perception in the domain of crosstalk cancellation and audio envelopment with a limited amount of loudspeakers. ... Read more

We all know the possible consequences of global warming, rising temperatures, flooded cities and destroyed ecosystems. One of the causes is the emission of gases, predominantly CO2, which is increased by the growing E-commerce market. E-commerce companies rely on recommender systems to stimulate users to purchase products. We are convinced that we can use the core strength of recommender systems, influencing decision making, to steer users towards eco-friendly choices. Therefore, in this thesis, we research how greenness can be integrated into recommender systems. We present the first recommender system dataset that includes greenness, we benchmark several recommendation algorithms and we propose a strategy to increase recommendation greennness. To create the dataset, we annotate an existing recipe recommendation dataset with recipe greenness. For our benchmarking experiment, we propose metrics to measure recommendation greenness, which we use to show that no recommendation algorithm is fundamentally greener than others. Lastly, we propose a re-ranking method for improving the greenness of recommendation rankings. We use the method to explore the trade-off between accuracy and greenness and we show that it is possible improve the greenness of recommender systems significantly with little loss of accuracy. ... Read more

In our daily life people encounter many social interactions, for example in the supermarket, at work and in schools. Currently the most reliable way to find social interactions in groups, is to manually annotate the data. Manual annotation takes a lot of time and human resources and as the information stream goes faster and faster, the manual labour cannot keep up. Therefore an automated program to replace this is desirable. One method to automate this, is by looking at the proximity of people, which has been done by Dikker [1]. The results look promising, however they are sensitive to errors. The F1-score (equation 3) is only 0.625 and the precision (equation 5) is 0.696. This means that there are still a lot of false positives. With orientation data one can determine if the people in question are facing each other before they are assigned to be in social interaction. With a view frustum (cone shape) it can be determined who are facing at each other by checking if the cones overlap. This cone can take different sizes, making the view area bigger or smaller. Because the absolute rotation were only used, it was used to find out who cannot face each other. Using the proximity results as a baseline, the impossible orientations were removed. With the goal to reduce the amount of false positives. From the results it can be concluded that a small cone will perform worse compared to the baseline. The bigger the angle of the cone becomes, the closer it gets to the results from Dikker [1]. However, in none of the results it turns out that using absolute orientations improves the current work significantly. ... Read more