Open Access

This thesis tackles a common bottleneck in data-science courses: students struggle to turn a broad interest into a focused, workable project idea. This thesis set out to design and evaluate a compact assistant—EduIDEAtor—that makes this first mile simpler and more intentional. The tool uses a text-first interface with plain inputs, a small set of clearly different directions, and quick, reversible edits so students can steer ideas without losing momentum. After building and iterating the web appli- cation, The thesis evaluated how students experienced it and how it compared with familiar, non-AI brainstorming. The findings are consistent: navigation and input clarity were strong; students felt more able to generate and shape ideas; overall sat- isfaction and willingness to continue using the tool were high. Two practical refine- ments emerged—make back navigation clearly visible and give users finer control over how broad or specific the suggestions are both achievable without changing the core design. The contribution is a concrete pattern for first-mile ideation and a set of actionable guidelines for course-level adoption.

The integration of Large Language Models (LLMs) into information retrieval sys tems has transformed the user experience by providing direct, conversational responses instead of traditional ranked lists of search results. This modification raises substantial concerns about user trust, behaviour, and the risk of misinformation, even as it improves accessibility and convenience. This thesis investigates the impact of generative information retrieval on the reliability of synthesized answers, particularly focusing on how hallucination rates and semantic drift influence trust dynamics and information-seeking behavior. By evaluating the performance of different LLMs on fact-checking benchmarks, the study seeks to quantify the advantages of model scaling against the inherent risks of factual inaccuracy. The study evaluates hallucination and user trust in LLM-augmented information retrieval systems using three fact-checking datasets. Three well-known semantic similarity metrics are employed to assess the alignment between LLM responses and ground-truth references. Furthermore, the hallucination rate and factual consistency are assessed by aligning model-generated responses with verified annotations in fact-checking datasets. We utilise bias detection measures to evaluate implicit stereotype reinforcement in LLM outputs. This study applies a comprehensive framework for evaluating and auditing hallucinations by combining quantitative performance metrics with user-level reliability insights. The work aims to establish a baseline for the transparency and reliability of LLMs in search and retrieval contexts.

Generative AI tutoring tools predominantly refine the wording, structure, and feasibility of a student’s initial problem framing. Whether AI assistance that instead poses counter-proposals – alternative framings, stakeholder perspectives, and questioned assumptions – changes the originality of the resulting framing is an open empirical question for data science education. This thesis investigates that Question in an exploratory study with thirty postgraduate students at Universität Koblenz. Each participant interacted with a bespoke web platform that walked them through three fixed-order stages on a sharedWi-Fi-dormitory dataset description: an Editorstyle AI stage, a Challenger-style AI stage, and a final unassisted synthesis. The final synthesis topic was rated on four dimensions (originality, feasibility, reasoning quality, clarity) by a human expert rater and an LLM second rater (Anthropic Claude) blinded to group assignment. Because the planned counterbalanced within-subjects crossover could not be implemented, participants were sorted post hoc into an Editor-influenced and a Challengerinfluenced group based on triangulated self-reported impact and preference. Topics from the Challenger-influenced group were rated substantially higher on originality (Cohen’s d = 3.67, 95% CI [2.50, 4.84], p < .001, ICC = 0.859); topics from the Editor-influenced group were rated higher on feasibility (d = −2.04, 95% CI [−2.92,−1.16]) and clarity (d = −1.07, 95% CI [−1.84,−0.30]); a secondary unexpected association favoured the Challenger-influenced group on reasoning Quality (d = 1.03, 95% CI [0.27, 1.79]). The latter three dimensions had inter-rater reliability below the pre-specified 0.70 threshold and their effect-size magnitudes are interpreted as exploratory estimates. Perception data and open-ended responses suggested participants treated the two AI styles as functional complements assigned to distinct task-contexts, and 94% indicated interest in an integrated switcher mode. Because the analytic groups were defined by self-report and because the stage order was fixed, these results should be read as associations rather than causal effects of prompt style. The thesis contributes (a) empirical evidence of large betweengroup differences in framing originality consistent with a counter-proposal mechanism; (b) a methodological case study of a hybrid human-LLM rating protocol with dimension-specific reliability documentation; and (c) design directions for educational AI tools that surface, rather than choose for the learner, the kind of cognitive assistance offered.

The proliferation of online abuse on social media platforms has emerged as a significant concern, negatively impacting users' mental health and online experiences. While the Natural Language Processing (NLP) community has developed various computational methods for abuse detection, including Support Vector Machines (SVMs) and Convolutional Neural Networks (CNNs), existing approaches predominantly focus on identifying explicit forms of abuse. This narrow focus overlooks subtle and contextual forms of online harassment, which can be equally damaging to users' wellbeing. This thesis presents a novel approach to online abuse detection by integrating contextual embeddings with sentiment analysis features through the fine-tuning of Large Language Models (LLMs). Our methodology leverages a comprehensive dataset of 47,000 annotated tweets for training, combined with sentiment analysis capabilities developed using 50,000 IMDB movie reviews. The system employs DistilBERT architecture to develop a sophisticated detection framework capable of identifying six distinct categories of abuse: ethnicity-based, age-based, gender-based, religion-based, other cyberbullying, and non-cyberbullying content. The author established a rigorous evaluation framework employing multiple metrics, including accuracy, recall, and F1 score, to assess the model's performance in detecting both explicit and nuanced forms of online abuse. The integrated system achieved an overall accuracy of 85\% across 6 categories on the cyberbullying dataset, outperforming other methodologies applied to the same data. In direct comparison, our approach— which uniquely combines contextual embeddings with sentiment analysis—demonstrated significant improvements over traditional fine-tuning methods, such as those using only BERT or RoBERTa, particularly in detecting subtle forms of abuse. Most notably, our system was more effective at identifying passive-aggressive content and context-dependent harassment, challenges that often cause conventional detection methods to fall short. This enhanced performance can be attributed to the model's ability to capture nuanced linguistic cues through its integrated analysis of both contextual information and sentiment, thereby offering a more refined interpretation of potentially harmful content. This research emphasizes the critical importance of incorporating subtle abuse detection into online content moderation systems. By developing more sophisticated detection methods that can identify both overt and nuanced forms of harassment, this work contributes to the creation of safer and more inclusive online spaces that facilitate constructive dialogue. The findings of this study have significant implications for the development of more effective content moderation tools and the broader goal of fostering healthier online communities.

Artificial intelligence (AI) is becoming more widely used in a number of industries, including in the field of education. Applications of artificial intelligence (AI) are becoming crucial for schools and universities, whether for automated evaluation, smart educational systems, individualized learning, or staff support. ChatGPT, anAI-based chatbot, offers coherent and helpful replies based on analyzing large volumes of data. Integrating ChatGPT, a sophisticated Natural Language Processing (NLP) tool developed by OpenAI, into higher education has sparked significant interest and debate. Since the technology is already adapted by many students and teachers, this study delves into analyzing the sentiments expressed on university websites regarding ChatGPT integration into education by creating a comprehensive sentiment analysis framework using Hierarchical Residual RSigELU Attention Network (HR-RAN). The proposed framework addresses several challenges in sentiment analysis, such as capturing fine-grained sentiment nuances, including contextual information, and handling complex language expressions in university review data. The methodology involves several steps, including data collection from various educational websites, blogs, and news platforms. The data is preprocessed to handle emoticons, URLs, and tags and then, detect and remove sarcastic text using the eXtreme Learning Hyperband Network (XLHN). Sentences are then grouped based on similarity and topics are modeled using the Non-negative Term-Document Matrix Factorization (NTDMF) approach. Features, such as lexico-semantic, lexico structural, and numerical features are extracted. Dependency parsing and coreference resolution are performed to analyze grammatical structures and understand semantic relationships. Word embedding uses the Word2Vec model to capture semantic relationships between words. The preprocessed text and extracted features are inputted into the HR-RAN classifier to categorize sentiments as positive, negative, or neutral. The sentiment analysis results indicate that 74.8% of the sentiments towards ChatGPT in higher education are neutral, 21.5% are positive, and only 3.7% are negative. This suggests a predominant neutrality among users, with a significant portion expressing positive views and a very small percentage holding negative opinions. Additionally, the analysis reveals regional variations, with Canada showing the highest number of sentiments, predominantly neutral, followed by Germany, the UK, and the USA. The sentiment analysis results are evaluated based on various metrics, such as accuracy, precision, recall, F-measure, and specificity. Results indicate that the proposed framework outperforms conventional sentiment analysis models. The HR-RAN technique achieved a precision of 98.98%, recall of 99.23%, F-measure of 99.10%, accuracy of 98.88%, and specificity of 98.31%. Additionally, word clouds are generated to visually represent the most common terms within positive, neutral, and negative sentiments, providing a clear and immediate understanding of the key themes in the data. These findings can inform educators, administrators, and developers about the benefits and challenges of integrating ChatGPT into educational settings, guiding improvements in educational practices and AI tool development.