The music listening preferences of children have been the subject of numerous studies, intending to inform the design of music recommender systems to better cater to children’s specific needs. Most of these studies are centered around genre, while few explore other traits of songs to capture listening behavior. In this research, we propose song structure, the arrangement of songs into sections, as a novel lens to analyze children’s music preferences. We create an extension from the popular LFM2b dataset to the Genius Lyrics dataset. Using this dataset, we group song interactions of users according to the educational level of children aged 12 through 17 and adults, and analyze and compare how song sections play a role in their listening behavior, based on the share of similarly structured songs in their listening data. To identify similar song structures, we cluster songs on their song structure fingerprints, extracted from their lyrics. We find no salient differences in the preferences of the children’s age groups, but we do find preliminary indications that adults interact less with songs characterized by the presence of lyrical Hooks (the most catchy parts of songs) than children and more with songs that are focused around Choruses. Although these insights can be used directly by recommender systems, our findings can be used as a springboard for future audio-based research to focus on not only lyrical, but also instrumental Hooks and their relation to children’s music preferences.

Music-streaming platforms rely on recommender systems to help listeners navigate millions of tracks, including a growing number of children using these platforms. However, most systems are optimized for adults, often resulting in recommendations that fail to reflect preferences or needs of children. While demographic features have been shown to improve performance in models focused on adults, their impact on a child-centric recommender system remains unexplored. This study investigates whether incorporating demographic features (age, gender, and country) and profiling features (exploratoryness, concentration, and replayness) improves the quality of music recommendations for children. Using a filtered subset of the LFM-2b dataset, we evaluate a baseline model based on implicit-feedback interactions against variants extended with different combinations of demographic and profiling features. Results show that demographic features led to a reduced accuracy across most models. In contrast, profiling features significantly increase top-K accuracy, with improvements up to 18%. These findings highlight the limitations of recommender systems tuned for adults when applied to children and emphasize the value of behavioral-aware modeling in the development of more effective child-centric music recommender systems.

Children are generally underrepresented in music recommender system research, despite having distinct preferences and developmental needs that set them apart from adult audiences. Traditional recommender approaches, designed primarily for adults, often fail to capture the unique listening behaviors of younger users and may fail to serve them effectively. At the same time, acoustic features play a significant role in shaping children's music preferences, yet their potential to enhance and provide optimal recommendations for children remains largely unexplored.

To address this gap, our study examines whether extending a standard collaborative filtering recommender with individual acoustic features can yield more age-appropriate music suggestions for children. We integrate content-based attributes, such as acousticness, danceability, energy, instrumentalness, liveness, loudness, mode, speechiness, tempo, and valence, into an item-based collaborative filtering algorithm and evaluate its performance on users aged 15 through 18. By comparing accuracy and
diversity metrics before and after the inclusion of each feature, we identify which acoustic feature improves recommendation quality for each age group and results in the highest performance.

Our findings emphasize the significance of acoustic features, including mode, instrumentalness, and acousticness, in improving performance metrics for distinct age groups. By identifying these age-specific features, our research contributes to the development of age-appropriate and child-centric music recommender systems.

Large language models (LLMs) are increasingly used by children, yet their responses are often not tailored to young users’ reading levels or cognitive development. Previous attempts to improve content readability through prompt modifications such as adding "for kids" have shown limited success. This project explores an alternative strategy: persona-based prompting. Rather than directly specifying the target audience, we instruct the LLM to role-play a teacher as a familiar figure to children. Using real child authored queries, we evaluate whether this role-based approach leads to more readable and comprehensible responses across different LLMs. Readability and comprehension were measured using established metrics, including Flesch-Kincaid formulas and Age of Acquisition data. Our results show that for 4 out of the four evaluated models, persona based prompting consistently produces responses that are more readable and accessible across all readability metrics and some comprehensibility metrics compared to standard or intended-user prompting. This finding suggests that persona-based prompting is a promising strategy for improving the suitability of LLM outputs for young audiences.

Children often struggle to retrieve age-appropriate information when seeking information online. One big reason for this is that their search queries are short, misspelled, or vague. As a solution to this problem, previous research investigated query reformulation, where the input query is manipulated in a way that the retrieved web results are more child-appropriate. This was measured by various metrics and scores, such as readability and content safety of the retrieved web search results. The problem with present query reformulation strategies, however, is that each tackles this problem from one perspective, missing out on the potential benefits of other perspectives. For instance, expanding the query with the “for kids” cue has shown to be a good way to target a specific audience and helps retrieve more child-appropriate content; however, on top of this considering substitutes for uncommon words with simpler synonyms might further enhance the child-appropriateness of the retrieved results as it tackles the reformulation from a different perspective than “for kids” audience cue expansion.

Motivated by this, we propose a multi-step query reformulation strategy that combines multiple reformulation strategies and applies them to the given input child query in a multi-step manner using a Large Language Model (LLM). We use LLM to apply the reformulation strategies to the input query in a chain-of-calls (where each call is prompted to apply a different reformulation strategy). This proposed method captures the perspective of multiple reformulation strategies, rather than a single one, unlike existing reformulation strategies. The results of our experiments, which include a baseline comparison (of the retrieved search results from the reformulated query against the original query) and an ablation study, provide insight into the performance of our strategy.

With this work, we aim to demonstrate the potential of combining multiple reformulation strategies and their impact on improving the readability and content safety of retrieved web search results when applied to children’s search queries. Our findings reveal a significant improvement in the readability of retrieved results after using the proposed reformulation method. Ultimately, this work contributes to the development of next-generation, child-centric search systems that deliver clearer, safer results for children.

Recommender Systems are key instruments that are constantly being employed in the online environment as a method of connecting users and items. Due to the resulting personalised suggestions, users benefit from quickened decision making, however, such systems can also introduce unwanted side-effects, by reinforcing existing biases or limiting exposure to diverse content. While much of the research surrounding unfairness and its effects has been conducted specifically to benefit the users, the creators of the distributed items, namely the providers, can also be subject to inequity. From the perspective of a provider, garnering visibility and exposure of their items through these systems converts into revenue. Recommender Systems can be a positive means in sectors where, historically, groups of providers have been disadvantaged from reaching their desired consumers, although if mishandled can become an additional hindrance. Amongst the many relevant domains, the book industry is a clear example where authors have been discriminated based on traits unrelated to the quality of their writing. Publishers have manifested an adversity towards translated works leading to an innate disadvantage when it comes to their distribution and marketing. Still, the fairness of how recommender systems handle translated works when competing with other books remains unexplored. To address this gap, we conduct an empirical exploration of several state of the art recommendation algorithms, evaluating their performance with respects to accuracy and provider fairness. This allows us to discern if any of the algorithms are a helpful conduit or a harmful one. Upon identifying inequity concerning foreign authors, we probe the ability of mitigating it on an algorithmic level, by applying a reranking method. Outcomes stemming from this work, reveal high representation of books available in English in recommendations, while highlighting the advantages of foreign manuscripts with translations over those without. The outlined findings inform future design of Recommender Systems and their provider fairness evaluation.

As online video platforms like YouTube and YouTube Kids continue to shape young children's daily media use, concerns about their exposure to inappropriate content persist. While these platforms implement various safeguards to protect young audiences, inappropriate videos continue to surface in search results and next-video recommendations, sometimes even at the top of the list. This study explores how metadata-derived features can be used to identify and address such content. Drawing on a ground-truth dataset of YouTube videos labeled for toddler appropriateness, we conduct a detailed feature analysis to uncover patterns linked to (in)appropriateness and train a classifier capable of predicting video appropriateness for young children. Building on these insights, we develop and evaluate score-based reranking strategies designed to reduce exposure to inappropriate videos while promoting age-appropriate content. Our findings show that metadata-informed reranking significantly improves the prioritization of suitable content, raising HitRate@1 of suitable videos from 14% to as high as 62%, but also reveal critical trade-offs: misclassified inappropriate videos, particularly when predicted with high confidence, may still appear in top-ranked positions. As such, detection and reranking methods like ours represent a first step that warrants further steps for safer recommendation environments for young children. This research provides a practical framework for improving recommendation outcomes and contributes to the broader conversation on designing safer, more transparent, and child-centered media systems.

Full Codebase available on the following repository - https://github.com/JoeydeW/KeepItPGorLetItGo

Autism Spectrum Disorder (ASD) is a neuro-developmental disorder reported to affect around 58 million people across the world. Due to their struggles with social communication in real life, they prefer to use online mediums to form social connections and are better able to maintain meaningful relationships through their communications over these online mediums. However, empirical studies have revealed that autistic users are not as efficient in using these online mediums when they have to search for information. While studies in the past have resulted in a growing list of guidelines to improve the web search experience of autistic users, whether existing online ISTs adhere to these guidelines has not been studied extensively. Thus while we know what the autistic users need to efficiently look for information on the web, we do not know if and how the ISTs are catering to their accessibility needs.

To advance knowledge in the accessibility of popular online ISTs when catering to the information needs of autistic users, we conducted an empirical exploration focusing on the accessibility of the textual
information in the IST responses. We specifically examined the responses on three aspects of text accessibility for autistic users - (1) text structure, (2) text readability, and (3) text concreteness using
“accessibility indicators” i.e. quantifiable features created by us based on widely accepted web content accessibility guidelines for autistic users. Due to the lack of standard query datasets for autistic users, we had to generate synthetic queries representative of autistic users’ information needs to collect responses from four popular ISTs Google, Bing, Gemini, and ChatGPT. We examined the differences in the way ISTs respond to queries characteristic to autistic users in contrast to queries typically asked by the general public, as well as the differences in IST responses across different kinds of queries asked by autistic users. We also investigated if an autistic user were to reformulate their query and explicitly state that they are autistic in their query would nudge the IST to respond differently to their query.

Juxtaposing the IST responses generated for control and ASD group queries revealed that the ISTs not only responded differently to the queries asked by the two user groups, but the responses generated for the control group were surprisingly more accessible for autistic users than the IST responses generated for their queries. For queries typically asked by autistic users, none of the ISTs produced responses
suitable for autistic users. However, each IST had its strengths in the context of the accessibility which can employed in the other ISTs to improve the general accessibility of the responses generated by all ISTs. We also found that our query and prompt reformulation strategies affected the accessibility of each IST differently. In general, the reformulations did nudge the ISTs to generate responses that are more suitable for autistic users. This presents an optimistic solution to the issue of ISTs not catering to the accessibility needs of autistic users when they search for information on the web. Our results highlight areas where popular ISTs can be improved to make their responses more accessible to autistic users. Our systematic empirical investigation pipeline can also be extended to investigate IST responses on other factors that influence the accessibility of web content for autistic users.

This thesis investigates the enhancement of sentence decomposition in Large Language Models (LLMs) through the integration of linguistic features, including constituency parsing, dependency parsing, and abstract meaning representation. Traditional decomposition methods, which often rely on rule-based approaches, struggle with highly intricate sentences. By incorporating detailed linguistic features and employing reasoning steps and supervision in prompts, we aim to improve the comprehension and decomposition capabilities of LLMs. Experimental results show that integrating these linguistic features has the potential to improve decomposition performance. However, these enhancements also introduce new error types and increase computational costs. This study identifies several error types in the programs generated by LLMs, including format, decomposition, and conversion errors, emphasizing the need for further refinement in model training and prompt design. This research lays a step toward more accurate and efficient processing of complex sentences in LLMs, encouraging ongoing development and optimization in this field.

While most large language models (LLMs) are powerful, they are primarily designed for general purposes. Consequently, many enterprises and institutions have now focused on developing domain-specific models. In the realm of education, an expert LLM can significantly enhance students' ability to find information more effectively and reach their learning goals. Nevertheless, the training of such expert models in education remains largely unexplored. This study explores this research gap by developing a framework to transform semi-structured educational web data into structured datasets and perform instruction tuning on foundation models. Additionally, we conduct a comprehensive performance analysis to determine how various training factors affect model performance.

We first employed a systematic and cost-effective approach involving web data extraction, data cleaning, validation, task design based on student surveys, and automated instruction instance generation using LLMs. Human evaluations confirmed the quality, especially the relevance and accuracy of these datasets.

This study then investigates the impact of various training techniques on domain-specific educational large language models (LLMs) performance. Our experiments reveal that further pre-training enhances model performance, especially with domain-specific terminology, although the performance gains decrease as the dataset size increases. Furthermore, multi-task training also improves model relevance, accuracy, and clarity, but less correlated tasks and datasets can present challenges. These challenges include increased complexity and potential degradation in performance due to the model having to switch between diverse tasks. Lastly, this study conducts a comparative analysis of different models and it highlights trade-offs between computational resources and performance.

The findings demonstrate that a structured approach to dataset generation and strategic training can effectively develop domain-specific LLMs in education. This research benefits the development of educational LLMs and provides a foundation for future researchers to build more specialized models in various domains.

Music recommender systems are increasingly present in our lives, and it is important to keep trying to improve recommendations in order to make them match the users preferences as well as possible. To achieve this, a vast amount of song and user data has to be analysed and taken into account. One of the approaches to do this, includes analyzing different audio features in order to find other songs with similar traits. The majority of the research and data in this sector is focused around adults, with little research surrounding children, which can result in worse recommendations for this demographic. In this paper, the focus is shifted towards children with the purpose of filling that gap. This is achieved by examining the prominence of specific song features among children of different age groups, expanding the knowledge on listening habits of a major demographic. More specifically, the research presented in this paper explores the prominence of various song features, aiming to find a connections between these features and the listening habits of children of specific age ranges from 8-18. This paper’s conclusions will offer potential enhancements, which can improve existing recommender systems by considering findings for their design. These findings will therefore allow for a more tailored experience for children of different age ranges, increasing overall user experience.

Recommender systems play a considerable role in the consumption of music, also for children. Children are easily influenced, inappropriate song lyrics can negatively impact children's behaviour and personality, by teaching inappropriate language or harmful biases. We argue for a need for recommenders to protect children from inappropriate content. Recommender systems should consider what content is and is not well-suited for children. To guide future design and development of recommender systems for children, this research probes the music lyrics of Pop, Rock, Country, Rap, and R&B for inappropriate content. We achieve this by an empirical analysis using four existing algorithms to scan different facets of inappropriate content in the lyrics in a dataset of 37,993 songs from the Genius database. The outcome of this empirical exploration reveals that Rap and R&B are the most inappropriate music genres and advises to be cautious when recommending these genres to children. Results also show that inappropriate content is highly prevalent in music. We address the need for a filter for recommender systems, to filter inappropriate songs for children.

The exponential growth of online content and consumer options has increased the reliance on recommender systems. Children, as a distinct user group, require tailored recommender systems different from those for adults. However, research on recommendation models for children is limited. This study evaluates deep learning recommendation models according to several performance and beyond-performance metrics on data from underage users on the Last.fm streaming platform, offering insights into optimal recommendation strategies for this demographic. Traditional non-deep learning models are used as baselines.

The online presence of e-commerce platforms, publishers, and other businesses has been increasing with the rise of internet use. These companies regularly expose their (potential) customers to massive amounts of media and products. It is vital for these companies to elicit interactions from these individuals in order to sell products or subscriptions. Engagement with online items, such as articles or products, are often measured using clicks. A well-known and proven method for improving online interactions is the use of clickbait. However, this practice is known to degrade the reputation of businesses. Widespread use of obvious clickbait is, therefore, not a sustainable solution for reputable companies. Research into alternative solutions to foster engagement is mostly taken from the perspective of the consumer. These solutions mainly involve various recommender algorithms that match consumers with items and articles that interest them. However, research from the perspective of the producer is extremely limited. The producer of the content has different incentives than the consumer of the content. Their main goal is to elicit an interaction from the consumer, while the consumers' main goal is to find something that interests them.

In this thesis, inspired by the work on personalization for consumers, we make the producer the main stakeholder. We explore whether the personalization of subject lines for groups of users with similar behaviour patterns and interests improves the engagement of consumers on newsletters. With the digital marketing company Basedriver as the use case, we personalized the headlines of marketing e-mails in newsletters and performed experiments to analyse the effectiveness of our technique. We hypothesise that e-mail marketeers can improve engagement by personalizing subject lines to address the personal interests of consumers. By looking at the content of articles previously clicked by consumers, they can be grouped into user groups with similar interests. Information on the interests of users from these groups can then be leveraged to generate a subject line addressing the interests of the group. The historic customer behaviour data and platform used for the validation of our new personalization strategy was provided by Basedriver and Hearst, who assisted in the research. Basedriver also provides the platform for collecting the data from the experiment.

This thesis presents the design and evaluation of a comprehensive system for developing voice-based interfaces to support users in supermarkets. These interfaces enable customers to convey their needs across both generic and specific queries. While current state-of-the-art systems like GPTs by OpenAI are easily accessible and adaptable, featuring low-code deployment with options for functional integration, they still face challenges such as increased response times and limitations in strategic control for tailored use-case and cost optimisation. Motivated by the goal of crafting inclusive, personalised, and efficient conversational agents, this study advances on three fronts: 1) a comparative analysis of four popular off-the-shelf speech recognition technologies to identify the most accurate model for different genders (male/female) and languages (English/Dutch); 2) an assessment of the effects of personalised recommendations versus generic responses, using a blindfolded, counterbalanced within-subject experiment; and 3) the development and evaluation of a novel multi-LLM supermarket chatbot framework, comparing its performance with a specialized GPT model powered by the GPT-4 Turbo, using the Artificial Social Agent Questionnaire (ASAQ) in a counterbalanced within-subjects experiment and qualitative participant feedback. Our find-ings reveal that OpenAI’s Whisper leads in speech recognition accuracy across genders and languages, users significantly prefer personalised chatbots over the non-personalised counterparts and that our proposed multi-LLM chatbot architecture outperformed the benchmarked GPT model across all 13 measured criteria, including statistically significant improvements in four key areas: performance, user satisfaction, user-agent partnership, and self-image enhancement. The thesis concludes by presenting a simple method for supermarket robot navigation by mapping the final chatbot response to correct shelf numbers towards which the robot can plan sequential visits. This later enables effective use of low-level perception, motion planning, and control capabilities for product retrieval and collection. We hope this work encourages more efforts into using multiple, specialised smaller models instead of always relying on a single powerful model.

Scratch is a block-based programming language. It is designed to be simple and syntax error avoidant. This makes Scratch an accessible platform for cultivating coding skills. Many young learners are taught about different programming skills using various project types as examples. For instance, games are used as an engagement tool, and various games can motivate new learners to make their own. This influx of new published projects are manually classified into different types by using tags in the project descriptions. However, this manual classification only happens when the user adds the appropriate tags. This calls for an extension of that feature, which could help classify all projects published on the Scratch website. This has the goal of improving the browsing process, especially for new projects. To address this gap, in this initial iteration we leverage the fact that there are likely similarities, or even various project type defining features that would help improve the accuracy of classification through machine learning. Filling this gap also opens the possibility of automatic classification, depending on the accuracy of the results. Within this study, various machine learning models were tested with quantifiable project features as input data. The accuracy scores were compared to draw conclusions on how well various features extracted from Scratch projects performed for classification.

Although existing work demonstrates that the usage of Scratch strengthens programming skills, little is known about the effect of children encouraging or criticizing each other when programming on the development of their programming skills. To address this gap, we conducted a data analysis to reveal the correlation (or lack thereof) between the sentiment of comments on a Scratch project and the creator's inclination to continue producing projects.

In particular, using a known dataset of Scratch projects, which we enriched with new projects, we examined several independent variables that capture sentiment on comments from different perspectives, e.g. the absolute or relative number of comments with a particular sentiment score that a user receives, or the ratio of comments that fall in a certain category. We also examined several dependent variables that capture the creator's inclination to continue producing projects, e.g. the number of projects they create or the total time that they are active on the platform.

The results of our experiments reveal that the absolute and relative number of comments with a specific sentiment score given to a user have a low correlation to the time that a user is active on the platform. Additionally, the ratio of the number of positive and negative comments over the total number of comments given to a user has a low correlation to the time that a user is active on the platform. Finally, the number of comments that a user receives has a low correlation to the number of projects a user creates.

Scratch is a popular, visual programming language aimed at children, and is used by teachers and after school code clubs to teach their students about programming. Measuring whether they understand the underlying concepts, however, is a difficult task. In this research, we tried clustering Scratch projects by complexity to help students improve their programming skills. We did this by selecting an existing data set to extract features that indicate code complexity. Before, researchers attempted clustering on one metric that globalises the project’s complexity. Different researchers set out to measure the growth of the students by clustering the projects the students created.
With that in mind, we adopt a partition-based clustering algorithm to cluster the projects, as this method indicates outliers. We examine the quality of these clusters using the silhouette coefficient. We set up five experiments with different input vectors to make out the impact each input has on the clusters. We did not find a clear indication of the projects being clustered by the selected features. This could mean that Scratch projects are not suitable to measure a high-level understanding of programming concepts. Including the project name in the input vector had a negligible effect on the outcome of the experiments.

Children spend a significant amount of time listening to music. Music has a significant cognitive and developmental effect on them. Because of their unique behavioral characteristics and their emotion regulation skills, their music preferences differ significantly from adults. However, limited study has been conducted on their music preferences and no music recommendation system have been designed to cater for their specific needs and preferences. With this study, we conduct an empirical exploration of the music preferences of children in terms of audio and sentiment characteristics, readability and topics discussed in the songs listened by children at different ages. We utilize the outcomes of our empirical exploration to adapt a recommender system to cater for the music preferences of children. Both the empirical exploration and the evaluation of the recommender system is based on the well-known extensive music dataset LastFM-2b. Outcomes from this work showcase that grade school students prefer more positive and joyful sentiments expressed in the lyrics, simpler language and higher acousticness. Older children prefer songs that convey sadness or anger, higher language complexity and songs they can dance to. Incorporating the song features into a recommender system leads to an increase in terms of all evaluation metrics in children compared to the RS trained only on user-item interaction data.

The spread of fake news has negatively impacted society. Prior efforts in Natural Language Processing (NLP) have employed machine learning models and Pre-trained Language Models (PLMs) like BERT to automate fake news detection with promising results. These models excel at text classification tasks, but their dependence on large context-specific datasets presents a hurdle in the dynamic and ever-evolving context of fake news. The recent emergence of Large Language Models (LLMs) offers a potentially transformative innovation. LLMs have demonstrated great promise in NLP tasks with little additional training data. Compared to PLMs, LLMs have broader knowledge and enhanced reasoning capabilities, suggesting their suitability for fake news detection. This study proposes an investigation which considers multiple perspectives to probe the applicability of the effectiveness, opportunities and challenges associated with leveraging LLMs for automated fake news detection. We leverage multiple state-of-the-art LMMs by using them at multiple levels of guidance to validate their accuracy and task-specific bias. The main contribution of this work is a better understanding of the role LLMs can play in automated fake news detection, the pitfalls that should be avoided when leveraging them, the most effective approaches when employing them, and the future challenges that need to be addressed. Our key contributions include a better understanding of how to employ LLMs for fake news detection, their strengths and weaknesses, and some practical recommendations for deploying them in the real world. Our work shows that while LLMs hold great potential for advancing automated fake news detection, thoughtful consideration of their limitations and careful application refinement are essential for their effective deployment in the fight against fake news.