05 Fakultät Informatik, Elektrotechnik und Informationstechnik
Permanent URI for this collectionhttps://elib.uni-stuttgart.de/handle/11682/6
Browse
208 results
Search Results
Item Open Access Prompt-based personality profiling : filtering social media posts using reinforcement learning(2024) Hofmann, JanAuthor profiling is the task of inferring characteristics about individuals by analyzing content they share. To date, systems that perform this task automatically, predominantly use supervised machine learning approaches and borrow from advances in the field of natural language understanding. However, while for many language understanding tasks immense progress has been made in recent years (e.g. by using pre-train then fine-tune paradigm), such progress most often does not transfer to automated profiling systems directly. One reason for this is that author profiling is inherently different from typical text inference tasks due to the possibly large amounts of content associated with an author. Therefore, this work proposes a new method for profiling that tries to select the most relevant parts of content shared by an author before inferring a characteristic. Here, instead of relying on ground-truth labels, this work uses the feedback from the zero-shot capabilities of a large language model to learn such a selection model via reinforcement learning, and evaluates this approach for personality profiling in social media. In experiments predicting big five personality traits, this work finds that prediction quality of such a system is comparable yet slightly worse to using all content associated to a profile in a zero-shot setting, while prediction time is reduced significantly due to the limited amount of content used for inferring personality in the proposed method. In addition, this work finds that simply selecting content arbitrarily leads to performance degradation for most traits, and therefore, this work concludes that, to some extent, the proposed method is able to distinguish between relevant and irrelevant content. Further, this work compares the proposed approach to existing supervised approaches and finds that such methods outperform the proposed method substantially. Still, since the ability of the proposed system to distinguish between relevant and irrelevant content of authors is closely tied to the capabilities of large language models, it can be expected that, with advances of such models, prediction quality of the proposed approach will increase in the future.Item Open Access Enhancing HTN planning with deep reinforcement learning for method selection(2025) Bahrami, SepidehAutomated planning is a central area within Artificial Intelligence (AI), enabling intelligent behavior in domains such as cloud computing, autonomous systems, context-aware activity recognition, and smart environments. Hierarchical Task Network (HTN) planning, which decomposes complex tasks into simpler subtasks using predefined methods, has proven effective in such structured domains. However, its performance is often constrained by static method selection strategies that lack adaptability to varying planning contexts. To address this limitation, this thesis proposes a neuro-symbolic framework that integrates HTN planning with Deep Reinforcement Learning (DRL), combining the strengths of symbolic reasoning and data-driven learning. Among the available DRL algorithms, Deep Q-Learning (DQL) is particularly suitable due to its off-policy nature, batch-efficient learning, and robust generalization across symbolic planning states. These characteristics align well with deterministic and hierarchical planners, enabling offline learning from curated datasets without requiring interactive exploration. The proposed integration introduces a learning-based decision layer that improves adaptability while preserving the reproducibility and determinism of the underlying planner. The effectiveness of this approach is demonstrated through a comprehensive evaluation across planning efficiency, memory consumption, and plan quality. Results highlight the potential of reinforcement learning to enhance classical HTN systems and support intelligent decision-making in complex, structured environments.Item Open Access Enhancing automotive safety through an ADAS violation dashboard(2024) Senger, TobiasAutonomous Driving (AD) is an active area of research in which Advanved Driver Assistance Systems (ADAS) play an important role. Ensuring the safety of ADAS systems is critical. However, most ADAS systems nowadays make use of Deep Learning or other types of Machine Learning. Formally verifying these systems to ensure their safety is hardly possible. For this reason, Radic explored the use of Runtime Monitoring (RM) to ensure the safety of ADAS systems by detecting violations of several specified Safety Requirements (SR) at runtime. After performing a test run with the system, she manually analyzed the causes of each series of violations in the extracted Violations Report. As this was laborious and time-consuming, this thesis should explore available approaches and techniques to automatically derive the root causes of violation series. To do this, we first perform an exploratory literature search. This allows us to identify that the most suitable approach to address our problem is Root Cause Analysis (RCA) using Language Models (LMs), Large Language Models (LLMs), Knowledge Graphs (KGs), or a combination of them. We perform a Rapid Review (RR) to find concrete techniques for this approach. We then conduct a narrative data synthesis to explore the techniques retrieved with our RR. This allows us to derive a plan to automatically analyze the causes of SR violations in a Violations Report. Our solution is then incorporated into a web-based safety dashboard application. This application enables our safety engineers to configure ADAS use cases, test tracks, and test runs. Then, the safety engineer can select a test run to display an interactive view of the test run. The safety engineer can then select individual violation series and analyze their root causes using our automated RCA solution based on LLMs. To evaluate the effectiveness of our system, we conduct a simple experiment. This experiment shows that our system already achieves comparable performance to a human baseline provided by Radic. Our system, therefore, represents a valuable tool for safety engineers to identify and repair safety-critical problems in ADAS systems in the context of AD. We also propose modified variants of our system that allow researchers to improve our automated RCA system in the future, e.g., by incorporating a KG.Item Open Access Stationary vehicle classification based on scene understanding(2024) Wang, WeitianNavigating through dense traffic situations like merging onto highways and making unprotected left turns remains a challenge for the existing autonomous driving system. Classifying vehicles into parked, stopped, and moving vehicles can benefit the decision-making system in this case because they play different roles during the vehicle-to-vehicle negotiation process. Existing works in vehicle classification focused on trivial cases and used methods that are not generalized enough. To fill this gap, after analyzing this problem and summarizing the necessary information needed for this problem, we propose a multi-modal model that can leverage information from lidar, radar, camera, and high-definition maps. To meet the complexity of our task and the needs of our model, we collect the dataset in real driving scenario and then preprocess and label it. By utilizing a pretrained vision encoder for fine-grained visual feature extraction and vision foundation model (CLIP) for scene understanding, our model achieves a 97.63% test accuracy on our dataset. Through visualization methods, experiments, and quantitative analyses, we investigate the effectiveness and importance of different encoders used in our model. We interpret and explain the successes and failures of our model to give a better understanding of how different latent features contribute to the final result. In the end, the limitations of our model and potential improvements are discussed.Item Open Access Cross-device data collection(2024) Zhu, KaiyangProviding cross-device service has become the best practice for many organizations. Those cross device services often have different user interfaces (UIs) designed to deliver user experiences (UXs) adapted to the device context. User behavior tends to vary across devices due to the UI variations, which can considerably impact usability. While extensive research exists on the relationship between user behavior and usability within the same device context, little attention has been paid to cross-device scenarios. By examining usage across devices, it becomes possible to identify the distinct user behavioral patterns and assess usability across varying UIs. Usability can be measured explicitly through user responses to a usability questionnaire and implicitly via navigation behavior events and task completion time on web pages. To facilitate this analysis, we developed a recording system capable of capturing user navigation behavior, usability feedback, and screenshots for five popular individual web pages. Participants were recruited both online and onsite to generate a diverse user behavior dataset. A total of 79 participants were invited to the experiment, with 17 participating in a pilot study. The resulting dataset comprises 34 well-processed, complete user datasets for analysis. Using this novel dataset, we conducted a preliminary analysis to explore correlations between user behavior features and usability metrics, specifically System Usability Scale (SUS) scores and task completion times. Additionally, we performed a small case study to compare the five different UIs and determine which design provided the best user experience. We found that usability, measured in terms of SUS scores, showed no significant differences across devices or web pages. However, completion time varied significantly across devices. Additionally, we provided design recommendations, such as addressing erratic user behaviors derived from speed-related metrics like excessive acceleration, which negatively impact usability, particularly on mobile devices.Item Open Access Tailoring language for user backgrounds : human and LLM evaluations of paraphrased texts(2025) Carlon, FrancescaIn an increasingly interdisciplinary world, the ability to tailor technical and field-specific language to diverse audiences has become essential. Both academic research and industry benefit from the knowledge and expertise of individuals from varied technical backgrounds, supporting progress in research and product development. Therefore, it has become essential to make domain-specific concepts accessible and understandable to different audiences. This study investigates how Large Language Models (LLMs) can generate effective paraphrases by tailoring language to specific target audiences. It also explores which linguistic elements enhance conceptual understanding and whether LLMs can produce high-quality synthetic data for evaluating text quality. This research specifically focuses on tailoring language from Linguistics, Computer Science, and Computational Linguistics texts for audiences with diverse academic backgrounds, using a variety of LLMs. Methods employed include Prompt Engineering, linguistic metrics, and evaluations conducted by both human judges and LLM-as-a-judge approaches. The experiments focus on assessing prompt quality, examining the relationship between human evaluations and linguistic features, and analyzing the correlation between human judgments and LLM-as-a-judge assessments. The results demonstrate that combining Role Prompting with Chain-of-Thought (CoT) prompting effectively produces high-quality paraphrased texts. Additionally, syntactic and lexical features play a central role in enhancing text accessibility and users’ understanding of unfamiliar concepts. Finally, while the correlation between LLM-as-a-judge evaluations and human judgments is promising, it varies across different user groups. Therefore, this research demonstrates the potential of Prompt Engineering techniques for paraphrasing and tailoring language to users with different backgrounds, explores the linguistic features that enhance text comprehension, and examines the extent to which LLM-as-a-judge correlates with human preferences and can support the automatic retrieval of high-quality data.Item Open Access Accented character entry in XR(2025) Pencheva, KristinaEntering accented characters on physical keyboards presents considerable challenges in computing environments, requiring users to memorize complex key sequences that disrupt typing flow, particularly in multilingual settings. While research on efficient accented character entry remains generally limited, this thesis explores Augmented Reality passthrough environments as a potential solution to these challenges. AR video passthrough provides advantages by preserving direct visual contact with physical keyboards and hands while enabling virtual interface augmentation. This strategy takes advantage of users’ existing familiarity with physical keyboards and maintains the tactile feedback that typists rely upon, while allowing for enhanced interaction through spatial positioning of virtual interface elements. This research employed a two-phase experimental approach using a Meta Quest Pro headset with integrated eye tracking. The interaction technique requires users to press Alt plus a base letter to display accent options, direct their gaze toward the desired accented character in the displayed toolbar, and confirm selection through various mechanisms. A pilot study evaluated three confirmation mechanisms: KEYPRESS CONFIRMATION (secondary keypress), KEY RELEASE CONFIRMATION (key release triggers selection), and EYE BLINK CONFIRMATION. Based on pilot results favoring KEYPRESS CONFIRMATION for its reliability and user preference, the main study used this eye-gaze selection with KEYPRESS CONFIRMATION strategy to investigate three spatial positioning strategies for accent selection toolbars: KEYBOARD-PROXIMATE (near physical keyboard), TEXT-FIELD (adjacent to input area), and GAZE-RESPONSIVE (appearing at user’s gazelocation). Results demonstrate significant performance advantages for GAZE-RESPONSIVE and TEXTFIELD positioning over KEYBOARD-PROXIMATE placement. Accent character typing speed improved substantially for both preferred strategies compared to KEYBOARD-PROXIMATE positioning. System usability scores were significantly higher for GAZE-RESPONSIVE and TEXT-FIELD strategies. GAZE-RESPONSIVE positioning received the strongest user preference, while KEYBOARD-PROXIMATE positioning was consistently rejected as least preferred. These findings provide empirical evidence for optimal toolbar placement strategies in AR text input systems and contribute to future development of multimodal interfaces combining physical keyboards with eye-gaze selection.Item Open Access Benchmarking pre-trained language models for schema-agnostic entity resolution(2024) Bothmann, JanData integration is a process in which data from different sources are brought together to create a unified picture of the data. A vital aspect of this integration is Entity Resolution, which tries to identify elements that correspond to the same entity across multiple datasets. The complexity of ER tasks can vary significantly, as data exhibits different characteristics and levels of structuredness, which can influence the difficulty of the task. In this thesis, we evaluate how current state-of-the-art Entity Resolution systems perform when dealing with semi-structured data. To do this, several semi-structured ER benchmarks covering data from various domains were created for evaluation. Additionally, to explore how different data characteristics or other influencing factors impact the performance of matching systems, we developed the Benchmark Creator. This tool allows us and other users to generate benchmarks where data exhibits specific characteristics that may influence the complexity of the ER task. We used Ditto, Sudowoodo and the GPT4o-mini model to evaluate performance on the newly created benchmarks. Our evaluation reveals that Ditto and the GPT4o-mini model can effectively perform schema-agnostic ER on semi-structured data.Item Open Access Prompt-based continual learning for visual question answering(2024) Ostertag, MagnusIn an ever-evolving world, Continual Learning (CL) strives to enable a costly trained model to learn new tasks without forgetting previously acquired knowledge. This work critically examines current CL benchmarks for Visual Question Answering (VQA), identifying significant shortcomings in the construction introducing bias. To address these issues, we propose a new CL-VQA benchmark based on GQA, designed to be incremental in both the language and the visual modality. Combined with learning it in one modality only, it can offer rich new diagnostics for a model. Additionally, we extend DualPrompt, a prompt-based CL method, DualPrompt, to the multi-modal domain. Using Dark Experience Replay as a baseline, we evaluate the performance against the new benchmark.Item Open Access Industry practices and challenges of using AI planning : an interview-based study(2024) Vashisth, DhananjayIn the rapidly evolving landscape of industrial applications, AI planning systems have emerged as critical tools for optimizing processes and decision-making. However, implementing and integrating these systems present significant challenges that can hinder their effectiveness. This thesis addresses the urgent need to understand the best practices and challenges involved in designing, integrating, and deploying AI planning systems in industrial settings. Without this understanding, industries risk inefficient implementation, leading to poor performance and resistance from end-users. This research employs a methodology that includes a literature review and interviews with industry professionals and researchers to identify common strategies and obstacles practitioners face. The study examines existing literature to uncover reported best practices and challenges in AI planning systems. Interviews provide additional perspectives, enriching the data collected and ensuring a thorough analysis. The findings reveal best practices, including the importance of cross-disciplinary collaboration, robust data management strategies, and iterative development processes. Additionally, recurring challenges such as integration complexities, scalability issues, and the need for continuous system evaluation are identified. These insights highlight critical areas for improvement and offer practical recommendations for enhancing the effectiveness of AI planning systems in industrial applications.