Umeå University's logo

As privacy concerns increase and regulation against tracking-based advertisingtightens, contextual advertising—which targets ads based on webpage content ratherthan personal data—offers a compelling alternative. The shift towards this alternativeform of ad targeting is gaining momentum thanks to advancements in artificialintelligence (AI), which significantly improve the ability to interpret and categorizeonline content. This thesis explores how AI can interpret online contexts and leveragethem for targeted, privacy-conscious marketing.A key contribution is the development of methods for extracting opinions from textand structuring them into “opinion units”, leveraging the power and versatility oflarge language models. Opinion units consist of concise, context-rich excerpts thatcapture individual opinions, paired with sentiment metadata. The proposed methodsdemonstrate high accuracy in opinion extraction and show promise for downstreamapplications. For instance, in opinion search and topic modeling of customer reviews,the compactness and distinctness of opinion units enhance retrieval precision andproduces more coherent and interpretable groupings of opinions. This enables theidentification of specific aspects driving customer satisfaction, providing insights forproduct development and targeted marketing.Marketing experiments conducted in this thesis reveal how media contexts influenceadvertising perceptions. The findings demonstrate that engaging content and thecredibility of website sources create a spillover effect, enhancing the effectiveness ofassociated ads. Regarding brand safety—ensuring ads do not appear in brand-damaging contexts—the results suggest that proximity to negative news articles aloneis not directly harmful. However, marketers face increased risks when the advertisedmessage is associated with the negative context. To mitigate these risks, AI tools canbe used to detect and avoid potentially unsafe online environments.Finally, the thesis offers guidance on AI-driven ad targeting by outlining the trade-offsbetween contextual and personalized strategies, as well as manual versus automatedmethods. The discussion considers key factors such as marketing objectives, dataavailability, and ethical considerations alongside regulatory requirements. Thefindings serve as a foundation for making well-informed, strategic choices in thefuture of advertising targeting.

Contextual language processing concerns how information from the surrounding media environment, such as accompanying texts or interaction signals, is represented and used in automated language processing systems. Incorporating contextual information enables models to derive interpretations that go beyond isolated textual analysis and capture aspects of meaning grounded in real-world situations. Despite substantial research in language modeling, effectively processing contexts remains an appealing challenge, particularly in balancing the expressive power, interpretability, and adaptivity of the model, and its alignment with human understanding.

This thesis investigates contextual language processing from both methodological and applied perspectives, focusing on two related domains: natural language processing and contextual advertising. On the methodological side, the thesis formalizes the grammatical inference of finite-state transducers as a sequential decision-making problem. By integrating structured transducer models with reinforcement learning, it demonstrates how contextual languages can be learned dynamically while retaining interpretability. These learning-based approaches offer a systematic way to adapt language transformations to shifting contexts. Such adaptivity can, for instance, lead to more effective decisions in automated advertising auctions. On the applied side, the thesis examines contextual communication in real-world, user-facing systems through empirical studies in contextual advertising. These studies reveal a gap between the computational level of contextual relevance that is often optimized by automatic metrics, and human-perceived relevance as experienced by end users. The findings show that increased algorithmic precision or model complexity does not necessarily translate into improved user experiences, and that factors such as transparency and trust play a central role in contextual effectiveness.

Through theoretical, empirical, and conceptual analyses, this thesis demonstrates that effective contextual language processing necessitates bridging multidisciplinary perspectives, including algorithmic and perceptual ones. It frames contextual language processing as an interpretable, adaptive, and user-centered interaction process. The insights discussed in this work offer implications not only for algorithmic optimization, but also for modern online advertising and other user-facing applications where contextual understanding, practical constraints, and privacy considerations are critical.