Thesis Repository

Abstract

Argumentation Frameworks are an approach of formalizing arguments and their interrelations in a graph structure. They can be used to draw conclusions from this modelling of knowledge. Since argumentation is an important part of human reasoning, these graph structures can be considered easily interpretable, what makes this technology an interesting explainable artificial intelligence method. Although this is not their main purpose, Quantitative Argumentation Frameworks can be used to solve classification problems by following a new approach. This approach is based on constructing them out of multilayer perceptrons (MLP), based on the work of Potyka. In this thesis we were motivated to construct Quantitative Argumentation Frameworks out of sparse MLPs. A swarm intelligence algorithm, namely Particle Swarm Optimization (PSO), was developed to search for sparse MLP models with specific characteristics that relate to performance and topology of the graphical structures. Models were implemented, tested, and evaluated on three different datasets. The implementation includes preprocessing of the datasets, parameter learning of MLPs based on backpropagation, and structure learning of the MLP graphical structures. The evaluation involves constructing fully connected MLPs and decision trees for comparison purposes. The resulting models achieved high performance and low complexity in their structure. The PSO algorithm also proved its efficiency in solving the structure learning.

Abstract

The extraction and segmentation of references from scientific articles is a core task of modern digital libraries. Once references are extracted and segmented, the bibliographic information can be made publicly available and linked, enabling efficient literature study. However, references often vary in their structure and content. This makes the extraction and segmentation of references a challenging but valuable task. The purpose of this thesis is to investigate whether Bidirectional Encoder Representations from Transformers (BERT) is suitable for the extraction and segmentation of bibliographic references. Therefore, we follow a deep learning approach for the extraction and segmentation of references from PDF documents. We use a neural network architecture based on BERT, a deep language representation model that has significantly increased performance on many natural language processing tasks. Over the BERT output, we put a linear-chain Conditional Random Field. We experiment with different BERT models and input formats and also examine two approaches for reference extraction and segmentation. The experiments are evaluated on a challenging dataset that contains both English and German social science publications with highly varying references. Our results show that the best performing BERT models were pre-trained on similar data to the data that we used for the fine-tuning of the BERT models on the task of reference extraction and reference segmentation. Moreover, our findings show that long, context-based input sequences yield the best results. The extraction model identifies and extracts references with an average F1-score of 81.9%. References are segmented with an average F1-score of 93.6%. We show that our models compare well to one other previously published work. Our conclusion is that BERT is a suitable choice for reference extraction and reference segmentation.

Abstract

Today, small satellites are an increasingly important way for researchers to launch experiments and payloads into Earth orbit. On the one hand, these satellite systems have a growing need for data transmission to a ground station; on the other hand, various constraints apply to the design and operation of radio communication systems in space. Today's radio systems usually have static transceiver configurations and the transmission rates used correspond to the worst-case consideration for the particular case. Recently, adaptive approaches for the operation of radio links between satellite and ground station have been described, in which the transceiver configuration, controlled by a learning algorithm, is changed during operation and adapted to the respective external influences. The basis for the algorithms used is the reinforcement learning model. The goal of this work was the further development of an algorithm developed for satellite communication and the testing of this algorithm within a simulation environment in order to evaluate the effects of the changes. The modified algorithm was run in a software simulation environment that included the digital signal processing of the transmitter and receiver and a model for the changing transmission conditions during a satellite overflight. The modification was compared with the original algorithm with respect to the achievable data transmission. Among other things, the number of neural networks used for learning was varied. Furthermore, different hyperparameters were varied and the effects on the data transmission were examined. An adjustment of the hyperparameters led to the transmission of 57.8% more data than in the baseline implementation of the algorithm. Both, the change in architecture and the reduction in the number of neural networks executed in parallel, led to slight performance losses of 4.0% and 3.3%, respectively. The results of the software simulations performed here, however, cannot be directly transferred to hardware simulation environments or actual communication links. The implementation of the different algorithm variants allows to test these variants with hardware transmission links in the future and to test such a system on a small satellite later on.

Abstract

A knowledge graph is a datastructure that is capable of storing knowledge. Besides that, there are several methods that use knowledge graphs to derive more information. These methods need to be validated with example knowledge graphs. However, real data might not be available or not contain desired properties. Thus, there are use cases that benefit from the generation of synthetic knowledge graphs. To define a synthetic knowledge graph, there is the need of a characterization that expresses how the synthetic data should look. In this thesis, I use Horn clauses for this characterization because of their good balance of expressiveness and complexity, their use in the field of rule mining, and their base role in the logical language Datalog. As clauses are usually not represented perfectly in real data, the goal of this thesis is to generate a knowledge graph that does not perfectly fulfil given Horn clauses, but in a desired degree of fulfillment. During the thesis, I developed and implemented two modifiable algorithms to generate knowledge graphs. On the one hand, I adapted the general hill climbing technique to generate knowledge graphs. On the other hand, I implemented a greedy algorithm which orders a given set of Horn clauses using logical subsumptions between their bodies, and then add edges to fulfil one Horn clause after the other, in the computed order. Both algorithms aim to fulfil the goal of this thesis by generating synthetic knowledge graphs according to given Horn clauses, each with a degree of fulfilment. The degree of fulfilment of any Horn clauses is characterized by body support, the number of times the premise of the Horn clauses is fulfilled, and support, the number of times the premise and conclusion of the Horn clauses are fulfilled. Additionally, there is the confidence which is the fraction of support and body support, i.e., the percentage of cases the Horn clause is fulfilled. All code is published such that anyone can try it. During the evaluation, random sets of Horn clauses were produced and the implementations generated corresponding knowledge graphs. Generated knowledge graphs were compared by considering the difference between the expected and the actual degree of fulfilment for each Horn clause. The result is that generation variant hill climbing with the initial state set to the result of the greedy algorithm with rule order based on subsumption yields the best results. Also, the difficulty of generating a good knowledge graph increases along with the overlapping degree of the input set of Horn clauses. Note that the overlapping degree reflects how many relation names occur in how many Horn clauses of the set. Lastly, the state-of-the-art mining tool AMIE found many Horn clauses in the generated graphs which were not intended by the set given to the generator.

Abstract

In the field of knowledge representation, Knowledge Graphs~(KGs) play a crucial role in organizing and representing vast amounts of information. However, they are inherently incomplete, as links can be missing from the graph. The task of predicting such missing links is known as link prediction. Link prediction is performed in different settings, namely transductive, where all entities and relation types are observed during training; inductive, where some entities are unseen; and fully-inductive, where some entities and relation types are unseen. A considerable amount of effort has been devoted to improving link prediction in the transductive setting, with Knowledge Graph Embedding~(KGE) emerging as a prominent approach. While research on inductive and fully-inductive link prediction has been conducted, existing methods rely on additional information beyond the set of training edges. To leverage the rich transductive literature for the two inductive settings, recent work in the form of BytE proposed using a BPE tokenizer to adapt transductive KGE to the inductive settings. However, it relies on a pre-constructed token vocabulary that may not be suitable for all knowledge graphs, and its evaluation is limited to benchmarks that contain only a few test edges with unseen elements. This thesis builds upon BytE by introducing purpose-built token vocabularies, modifying the process of transforming token embeddings into entity and relation embeddings, and subsequently performing a systematic evaluation on standard inductive benchmarks.

Abstract

In recent years, large-scale knowledge graphs have emerged, integrating data from various sources. Often, this data includes assertions about other assertions, establishing contexts in which these assertions hold. A recent enhancement to RDF, known as RDF-star, allows for statements about statements and is currently under consideration as a W3C standard. However, RDF-star lacks a defined semantics for such statements and lacks intrinsic mechanisms to operate on them. This thesis describes and implements a novel query language, termed eSPARQL, tailored for epistemic RDF-star metadata and grounded in four-valued logic. Our language builds on SPARQL-star, the query language for RDF-star, by incorporating an expanded FROM clause, called FROM BELIEF, designed to manage multiple, and occasionally conflicting, beliefs. eSPARQL’s capabilities are demonstrated through four example queries, showcasing its ability to (i) retrieve individual beliefs, (ii) aggregate beliefs, (iii) identify conflicts between individuals, and (iv) handle nested beliefs (beliefs about beliefs). The implementation of eSPARQL developed in this thesis is built on top of an existing SPARQL-star query engine. In this implementation, the execution process of a eSPARQL consists of two phases. First, the expression in the FROM BELIEF clause, called belief query, is translated into a SPARQL-star CONSTRUCT query that generates an intermediary graph, containing the beliefs of the subjects described in the belief query. In the second phase, This intermediary graph is then processed with the graph pattern of the eSPARQL by translating it to a graph pattern that can be processed by a standard SPARQ-star engine. In this last phase, the implementation translates eSPARQL operations to SPARQL-star, and checks if the pattern contains nested eSPARQL queries to be processed recursively. We study two research questions: (RQ1) Does the eSPARQL implementation scale? and (RQ2) How the eSPARQL implementation execution times compare with the execution time of manually written SPARQL-star queries? To answer these research questions, use the four example eSPARQL queries that showcase the abilities of eSPARQL and create a synthetic dataset generator that generates graphs of multiple sizes. Additionally, for research question RQ2, we manually generate SPARQL-star queries that are equivalent to the example eSPARQL queries. Regarding research question RQ1, our results show that eSPARQL has an execution time that is proportional with the data size. Regarding research question RQ2, except for one question, the manually written SPARQL-star queries are clearly faster than our implementation. Although the implementation showed to be slower than the manually generated SPARQL-star queries, the eSPARQL queries are shorter and easier to understand. This positive aspect of eSPARQL, can motivate further studies on how to optimize the eSPARQL implementation.

Abstract

Multistep prediction is the prediction of states based on initial states and a series of control inputs. This paper focuses on developing transformer models for multistep prediction of vehicle states and testing different modifications of the transformer architecture using the example of the prediction of a ship simulation. Research in NLP promises advantages with regard to training time and prediction accuracy for the transformer architecture compared to a state-of-the-art LSTM model. The author also investigates whether positional encodings are useful in this scenario and if a transformer model can learn the order of the inputs without positional encodings.

Abstract

Named entity recognition (NER) is a fundamental task in Natural Language Processing (NLP) that involves identifying and classifying entities in text. NER forms the basis for solving many NLP tasks that play an important role in the biomedical domain. However, the scarcity of annotated data poses significant challenges in training effective NER models, particularly in low-resource languages, such as German. This study aims to evaluate five models in a low-resource environment, such as oncology in German, to gain insights into possible solutions to overcome this data scarcity. These approaches include prompt engineering of a Large Language Model, a BERT-based model, fine-tuned separately on the WikiMed-DE-BEL and a synthetic dataset, and a GLiNER-based model architecture. A model fine-tuned on a German gold standard oncology dataset serves as a benchmark to compare the other four models to. The evaluation is conducted using two datasets: MAED-KS, a de-identified oncology dataset consisting of patient reports collected from a German hospital, and the aluminium standard dataset, a synthetic dataset constructed based on disease co-occurrence, extracted from hospital data. Our results show that the GLiNER architecture achieves the highest F1-scores of the four approaches on the MAED-KS dataset and similar scores to the benchmark model. The Prompt Engineered Large Language Model performed best on the aluminum standard dataset. Overall, the GLiNER model shows the best performance and good adaptability in a low resource environment, such as oncology data in German.

Abstract

Playing computer games is one of the most common and enjoyable activities in the modern lifestyle. However, people with motor impairments who cannot use mouse and keyboard are not able to interact with computing devices as required by the game design. In this regard, novel interaction techniques can enable hands-free control using gaze and voice as input modalities to assist people with motor and speech impairments. This is the first evaluation of a video game control method consisting of a combination of eye tracking and non-verbal voice commands (e.g. humming) applied in a 2D jump-and-run game environment involving essential spatio-temporal interactions. To assess interaction feasibility, the evaluation study consisted of both qualitative and quantitative measures. In addition, the feasibility was validated with a target user group of people with motor impairments. Overall, the results indicate a lower but competitive performance while increasing the fun factor.

Abstract

The modern world is highly influenced by user-contributed content such as user comments, which is one of the most popular forms of communication on social media. They help build a connection between content creators and content consumers, as well as a connection between users of a social platform, which makes them highly relevant for community interaction. However, researchers have not treated users' comment reading strategies in much detail. The existing accounts are limited to address solely an explicit comment reading approach, which is only able to track an active interaction, e.g. analyze attention by counting a number of clicks. This technique can not fully estimate the drawn attention, as 73% of people do not actively interact with comments. Eye-tracking plays a vital role in solving the task of analysing implicit attention, as it allows to estimate such crucial characteristics, as comment features, which drew the most attention, the amount of attention given or the order of seeing specific comments. The current research uses eye-tracking based attention analysis for investigating the phenomena of users' reading behaviour on the real YouTube interface. The present master thesis concentrates on analysing users' attention mechanisms and reading behaviour and finding a correlation between the comment features such as length, language, sparseness, comment position, number of likes, presence of replies, presence of video creator's like and authorised user label, which cannot be exhibited from a pure textual point of view. This work shows that number of likes and presence of answers contribute the most to the attention drawn by comments. The analysis revealed that the category of a video deeply influences the emotion and length of the popular comments: if people are looking for useful information, which applies to educational videos, they tend to pay attention to neutral and long comments, whereas while looking for entertainment, people most probably will notice short and positive comments. The two important findings from the gender analysis are that women tend to read longer comments, but skip a higher percentage of comments than men. It is hoped that this research will contribute to a deeper understanding of the features that draw the most attention, which can be exploited in content generation strategies and developing new ranking algorithms.

Abstract

The shift towards online meetings, seminars, and lectures has rapidly gained momentum in recent years, particularly accelerated by the events surrounding the COVID-19 pandemic. However, the absence of nonverbal cues such as eye contact and gestures in virtual settings presents significant challenges for effective communication. Existing frameworks for virtual meetings often fail to adequately address this issue, making it difficult for presenters to accurately assess audience engagement and comprehension. This thesis investigates these challenges by capturing participant attention and emotion in real-time and evaluating how presenters interact with live feedback. Through the implementation and study of a feedback system that visualizes gaze and emotional data, we aimed to bridge the gap between in-person and online lecture experiences. To ensure an informed approach, we conducted a requirement analysis prior to our experiment, identifying key factors for effective feedback integration. In the experiment, we collected data through our tool by logging meeting interactions and analyzing responses from a post-experiment questionnaire. Our findings provide insights into the impact of such feedback on presenter motivation, delivery adjustments, and engagement levels, ultimately contributing to the improvement of educational quality in virtual environments.

Abstract

Symbolic regression is an interpretable machine learning method that learns mathematical expressions from given data. It naturally combines with Bayesian Inference which lets experts express their knowledge as prior distributions over equations. However, the infinite search space of mathematical expressions renders exhaustive search impractical, and Bayesian Inference remains costly. Therefore, we propose to execute the Bayesian Reasoning in the learned latent space of a trained Variational Autoencoder (VAE) and thereby exploit inherent structures in the search space. While latent spaces have been used to structure search spaces, our approach provides the probability of each mathematical expression rather than selecting the best one. We suggest practical approximations to the posterior distribution in latent space and obtain formula examples by sampling from the posterior using the Gaussian Process Hamiltonian Monte Carlo (GP-HMC) method. We have validated our method using various Koza, Nguyen, and self-generated datasets and compared it against genetic programming and SInDy concerning the Root Mean Square Error (RMSE). Keywords: Symbolic Regression, latent space, Variational Autoencoder, Character Variational Autoencoder, Grammar Variational Autoencoder, Bayesian Reasoning, Gaussian Process, Hamiltonian Monte Carlo, Gaussian Process Hamiltonian Monte Carlo.

Abstract

Knowledge Graphs (KGs) are fundamental for organizing and representing large amounts of information, but they often suffer from incompleteness. Link prediction using Knowledge Graph Embedding (KGE) methods has emerged as a solution to this problem. Many different methods have been proposed to perform link prediction, some of which are a combination of different methods. However, existing approaches that combine different methods typically train models on the entire graph, lacking the diversity seen in machine learning ensembles such as bagging and random forests. In this thesis, we present the novel ensemble approaches UnifEnt and UnifFeat, that divide the KG into sub-graphs by taking advantage of the core principles of bagging and random forests. We evaluated our approach on common KG datasets and showed the benefits of using our method by comparing it to common KGE baseline methods, as well as related work in the area of ensemble methods for link prediction.