Content Analysis and Generation

This thesis tackles the connected challenges of content analysis and content generation through the development of explainable models for complex datasets. Current approaches often miss hidden information and fail to capture causal progression, limiting their ability to produce coherent, meaningful content. For image generation, creating effective prompts remain another bottleneck, which this work addresses using Large Language Models (LLMs) for prompt optimisation. These tasks are chosen because, together, they confront the core problem of bridging deep content understanding with the generation of high quality, explainable outputs. Accordingly, this thesis focuses on three main tasks: (i) Developing content models for identifying engaging features using Significant Feature Lasso (SFLASSO), (ii) Proposing content models for generating coherent story endings, and (iii) Designing a prompting approach for generating social media images. To support explainability, this thesis proposes content models based on SFLASSO, which identifies the most significant features influencing a target outcome. It also introduces a variant called SFLASSO-SI, which performs selective inference for a conservative significance test. Furthermore, a two-stage anchoring-and-adjustment approach is proposed: in the first stage, the optimal penalisation value is identified along the regularisation path; in the second stage, features with high Variance Inflation Factor (VIF) are removed to improve the quality of the selected features. For story generation, this thesis introduces content models that generate transitional commonsense sentences between adjacent context sentences, thereby revealing the hidden narrative. By modelling soft causal relations grounded in everyday commonsense knowledge, the approach enhances the causal coherence of generated stories. Causal reasoning and dependency parsing are used to capture long-range text dependencies, and multi-level Graph Convolutional Networks, along with LLMs, are employed to provide enriched contextual representations. Finally, the thesis presents a prompting approach for generating social media images. While prompt engineering has recently gained attention, most research focuses on text and reasoning. In contrast, this work demonstrates that prompt-aided image generation can outperform images designed by human experts. State-of-the-art models such as Midjourney, DALL·E 3, and Stable Diffusion XL are utilised to generate images from prompts.