Disinformation Detection: An Explainable Semi-Supervised Transfer Learning Approach

Disinformation, misinformation, and harmful online content have increasingly shaped public discourse, political decision-making, and societal trust. Through the expanding role of social networks and digital news platforms, users are confronted with large volumes of heterogeneous information, making it difficult to distinguish reliable content from deceptive or manipulative material. Existing machine learning approaches often focus on binary classification without addressing the linguistic, narrative, and contextual structures that contribute to misleading content. In addition, related forms of harmful communication - such as hate speech, toxic language, propaganda, biased reporting, and extremist narratives - are closely intertwined with disinformation, yet are rarely analyzed together within a unified research setting. To address these challenges, this dissertation presents DisDETECT, a system for the semi-automatic detection, analysis, and contextualization of disinformation in German-language online media. Two datasets were created for this purpose: one covering short texts and one consisting of long-form articles, both including disinformation, narrative structures, propaganda cues, hate and toxic language, and other stylistic or communicative phenomena. Based on these resources, an information extraction pipeline identifies entities, events, topics, claims, narratives, framing techniques, and indicators of bias at both document and corpus level. The extracted elements are integrated into a semantically enriched knowledge graph that organizes structured triples together with embedding-based representations for subsequent analytical steps and knowledge infusion. The methodological component of the thesis introduces a hybrid classification pipeline that combines transformer-based language models, linguistic features, and knowledge graph embeddings. Nineteen models were developed and evaluated across binary, multiclass, and multilabel tasks relevant to disinformation and harmful communication. Transfer learning was applied using multiple external datasets for pre-training and fine-tuning. The outputs of these models form the basis of an assessment mechanism that assigns degrees of reliability to textual content. To support interpretability, a rule-based explanation layer and large language model components generate textual explanations of model decisions, which are presented within the interactive visual analytics interface of the DisDETECT system. The results of this work show that the combination of information extraction, transfer learning, knowledge infusion, and hybrid classification methods supports detailed analysis of disinformation and related communication styles. A user evaluation of the DisDETECT system provides insights into how experts interact with the interface and interpret the presented results. By integrating linguistic, narrative, and contextual perspectives into a single analytical environment, this dissertation contributes to a broader understanding of how harmful content emerges, overlaps, and propagates across online media.