Year: 2026
Type: dissertation
Abstract: Social media has become indispensable in the digital age, frequently used for engaging with contemporary social agendas. These platforms not only shape public discourses but also influence individuals' perceptions and behaviours. This influence was especially evident during critical moments such as the COVID-19 pandemic, where social media played a central role in disseminating information and shaping public responses. As a result, key stakeholders, including policymakers, journalists, and communication professionals, increasingly rely on interpreting these dynamic social interactions to inform strategic decisions. This is due to the persuasive power embedded in social media, making it instrumental in guiding organisational objectives. However, the complexity and volume of social media discourses pose significant challenges for existing analytical tools in providing useful insights required for effective decision-making. Therefore, there is a growing need to enhance these analytical tools to support informed decision-making. Human-centred approaches to designing tools that integrate human expertise with the qualitative analyses of social media discourses remain underexplored. Addressing this gap requires understanding the limitations of both machine and human perspectives and how they can effectively complement each other. This thesis approaches the problem from two disciplinary perspectives. First, it applies Natural Language Processing (NLP) methods to establish the technical feasibility of analysing social media discourse and to identify where computational learning's limitations can be enhanced with human input. Second, it adopts Human-Computer Interaction (HCI) methods to design an interactive system informed by these technical findings, allowing design choices to be grounded in empirical evidence that highlights where and how human input can complement computational analysis. Then, the resulting interactive design is evaluated through user studies to examine how grounding the design in both computational limitations and human strengths influences human experiences during interaction with these automated analyses. For this purpose, the first part of this thesis experimented with a computational learning task, stance detection. This is a commonly chosen NLP method in social media analysis for misinformation because it involves identifying the attitude or opinions expressed on social media texts towards a target topic. Such positions reveal how narratives are spread and how discussions become divided or amplified, which are all common patterns in misinformation. The key challenge identified early in the study is the limited availability of datasets that capture the surrounding discourse to model contextual knowledge. To address this, the study adopted data augmentation, a widely used technique for mitigating data limitations in producing more relevant training samples. The study concluded by developing augmentation methods using advances in prompt engineering with Large Language Model (LLM) to generate synthetic conversations. The results demonstrated improved performance over baseline methods. While exploratory, the findings established the technical feasibility of modelling context learning through social media discourses and revealed limitations where the design of interactions can adopt the complementary human judgment. To bridge these computational insights with human expertise, the thesis introduces DiscourseDash, an interactive dashboard designed to support the exploration and interpretation of social media discourse. This design bridges the technical capabilities and human strengths as a combined effort to make more informed judgments by combining computational outputs with their interpretive skills. Guided by human-centred design principles, a functional prototype of DiscourseDash was developed to facilitate its evaluation through two phases of user studies. The first study explores how general users interact with DiscourseDash across a set of tasks designed to reflect real-world workflow scenarios. To test the hypothesis that DiscourseDash offers a more effective way of understanding social media discourse compared to navigating unstructured platforms, we examined participants' performance and experiences across two conditions. A mixed-methods approach was employed, integrating quantitative and qualitative data to assess the overall effectiveness of DiscourseDash and the usefulness of its proposed analytical features, stance, sentiment, popularity, and discourse-level insights in facilitating decision-making processes. The findings demonstrate both strong user preference for DiscourseDash and the perceived value of these analysis measures. The study also identified design limitations that outline opportunities for future improvements. To further assess its practical relevance, DiscourseDash was evaluated through a semi-structured interview study with industry experts. This aimed to explore how the tool aligns with established professional workflows and to capture perspectives more than those of general users. Experts recognised the tool's potential within its capacity to support discourse-level analysis and offer more qualitative insights. However, the study also revealed key misalignments between the design and expectations embedded in expert practices, which may affect adoption. Experts consistently emphasised the need for tools that move above surface-level statistical outputs to enhance support for the complexity and contextual sensitivity of social media discourses, strengthening the core motivation behind DiscourseDash's design. These two studies provided complementary perspectives. The first captures how non-expert users engage with the tool in tasks that simulate real-world decision-making scenarios. The second study contributes expert applicability by examining how DiscourseDash aligns with established professional practices. Finally, we discuss the broader implications of these findings for designing tools that focus on providing discourse-level insights. This thesis makes several key contributions by demonstrating how integrating discourse-level qualitative analysis with an interactive tool can enhance social media analysis for more informed decision-making. It introduces an interactive design approach through DiscourseDash, a dashboard that combines the strengths of computational methods with human judgment. Grounded in both technical findings and human-centred design principles, DiscourseDash emphasises contextual understanding to support decision-making workflows. The contributions include: (1) methodological reflections on the practical trade-offs and limitations of NLP methods addressing data limitation through augmentation and context modelling in a computational learning task; (2) the design and implementation of a functional prototype, DiscourseDash, informed by these reflections on computational learning capabilities and limitations, and (3) evidence from two user studies, one with general users and one with industry experts, that demonstrate the tool's effectiveness by showcasing its applicability across different expertise levels, and identify considerations for real-world adoption. Overall, this research emphasises the importance of ongoing investigation in discourse-level insights into analytical tools for social media analysis by demonstrating their relevance and applicability in real-world contexts. DiscourseDash represents a foundational step towards developing advanced interactive tools for social media analysis, built upon an empirical understanding of the practical workflows in which these tools are embedded.
Author: Liew, Xin Yu
Language: English
Subfield: General Social Sciences
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7218259497