Representation Learning techniques play a crucial role in a wide variety of Deep Learning applications. From Language Generation to Link Prediction on Graphs, learned numerical vector representations often build the foundation for numerous downstream tasks. In Natural Language Processing, word embeddings are contextualized and depend on their current context. This useful property reflects how words can have different meanings based on their neighboring words. In Knowledge Graph Embedding (KGE) approaches, static vector representations are still the dominant approach. While this is sufficient for applications where the underlying Knowledge Graph (KG) mainly stores static information, it becomes a disadvantage when dynamic entity behavior needs to be modelled. To address this issue, KGE approaches would need to model dynamic entities by incorporating situational and sequential context into the vector representations of entities. Analogous to contextualised word embeddings, this would allow entity embeddings to change depending on their history and current situational factors. Therefore, this thesis provides a description of how to transform static KGE approaches to contextualised dynamic approaches and how the specific characteristics of different dynamic scenarios are need to be taken into consideration. As a starting point, we conduct empirical studies that attempt to integrate sequential and situational context into static KG embeddings and investigate the limitations of the different approaches. In a second step, the identified limitations serve as guidance for developing a framework that enables KG embeddings to become truly dynamic, taking into account both the current situation and the past interactions of an entity. The two main contributions in this step are the introduction of the temporally contextualized Knowledge Graph formalism and the corresponding RETRA framework which realizes the contextualisation of entity embeddings. Finally, we demonstrate how situational contextualisation can be realized even in static environments, where all object entities are passive at all times. For this, we introduce a novel task that requires the combination of multiple context modalities and their integration with a KG based view on entity behavior.

This thesis explores local and global normal and Edgeworth approximations for symmetric binomial distributions. Further, it examines the normal approximation of convolution powers of continuous and discrete uniform distributions. We obtain the optimal constant in the local central limit theorem for symmetric binomial distributions and its analogs in higher-order Edgeworth approximation. Further, we offer a novel proof for the known optimal constant in the global central limit theorem for symmetric binomial distributions using Fourier inversion. We also consider the effect of simple continuity correction in the global central limit theorem for symmetric binomial distributions. Here, and in higher-order Edgeworth approximation, we found optimal constants and asymptotically sharp bounds on the approximation error. Furthermore, we prove asymptotically sharp bounds on the error in the local case of a relative normal approximation to symmetric binomial distributions. Additionally, we provide asymptotically sharp bounds on the approximation error in the local central limit theorem for convolution powers of continuous and discrete uniform distributions. Our methods include Fourier inversion formulae, explicit inequalities, and Edgeworth expansions, some of which may be of independent interest.

This dissertation focusses on research into the personality construct of action vs. state orientation. Derived from the Personality-Systems-Interaction Theory (PSI Theory), state orientation is defined as a low ability to self-regulate emotions and associated with many adverse consequences – especially under stress. Because of the high prevalence of state orientation, it is a very important topic to investigate factors that help state-oriented people to buffer these adverse consequences. Action orientation, in contrast, is defined as a high ability to self-regulate own emotions in a very specific way: through accessing the self. The present dissertation demonstrates this theme in five studies, using a total of N = 1251 participants with a wide age range, encompassing different populations (students, non-student population (people from the coaching and therapy sector), applying different operationalisations to investigate self-access as a mediator or an outcome variable. Furthermore, it is tested whether the popular technique of mindfulness - that is advertised as a potent remedy for bringing people closer to the self -really works for everybody. The findings show that the presumed remedy is rather harmful for state-oriented individuals. Finally, an attempt to ameliorate these alienating effects, the present dissertation attempts to find theory-driven, and easy-to-apply solution how mindfulness exercises can be adapted.

　

　

Environmental DNA (eDNA) metabarcoding promises to be a cost- and time-efficient monitoring tool to detect interactions of arthropods with plants. However, observation-based verification of the eDNA-derived data is still required to confirm the reliability of those detections, i.e., to verify whether the arthropods have previously interacted with the plant. Here, we conducted a comparative analysis of the performance of eDNA metabarcoding and video camera observations to detect arthropod communities associated with sunflowers (Helianthus annuus, L.). We compared the taxonomic composition, interaction type, and diversity by testing for an effect of arthropod interaction time and occupancy on successful taxon recovery by eDNA. We also tested if prewashing of the flowers successfully removed eDNA deposition from before the video camera recording, thus enabling a reset of the community for standardized monitoring. We find that eDNA and video camera observations recovered distinct communities, with about a quarter of the arthropod families overlapping. However, the overlapping taxa comprised ~90% of the interactions observed by the video camera. Interestingly, eDNA metabarcoding recovered more unique families than the video cameras, but approximately two-thirds of those unique observations were of rare species. The eDNA-derived families were biased toward plant sap-suckers, showing that such species may deposit more eDNA than, for example, transient pollinators. We also find that prewashing of the flower heads did not suffice to remove all eDNA traces, suggesting that eDNA on plants may be more temporally stable than previously thought. Our work highlights the great potential of eDNA as a tool to detect plant-arthropod interactions, particularly for specialized and frequently interacting taxa.

　

This meta-scientific dissertation comprises three research articles that investigated the reproducibility of psychological research. Specifically, they focused on the reproducibility of eye-tracking research on the one hand, and studying preregistration (i.e., the practice of publishing a study protocol before data collection or analysis) as one method to increase reproducibility on the other hand. In Article I, it was demonstrated that eye-tracking data quality is influenced by both the utilized eye-tracker and the specific task it is measuring. That is, distinct strengths and weaknesses were identified in three devices (Tobii Pro X3-120, GP3 HD, EyeLink 1000+) in an extensive test battery. Consequently, both the device and specific task should be considered when designing new studies. Meanwhile, Article II focused on the current perception of preregistration in the psychological research community and future directions for improving this practice. The survey showed that many researchers intended to preregister their research in the future and had overall positive attitudes toward preregistration. However, various obstacles were identified currently hindering preregistration, which should be addressed to increase its adoption. These findings were supplemented by Article III, which took a closer look at one preregistration-specific tool: the PRP-QUANT Template. In a simulation trial and a survey, the template demonstrated high usability and emerged as a valuable resource to support researchers in using preregistration. Future revisions of the template could help to further facilitate this open science practice. In this dissertation, the findings of the three articles are summarized and discussed regarding their implications and potential future steps that could be implemented to improve the reproducibility of psychological research.

Cette contribution pratique présente les principales méthodes d’analyse qui peuvent être utilisées dans diverses matières à l’école, mais qui peuvent aussi favoriser la réflexion sur des conflits personnels qui surviennent au sein de la famille, de la classe ou de la communauté scolaire.