THETA: A Textual Hybrid Embedding-based Topic Analysis Framework and AI Scientist Agent for Scalable Computational Social Science

Researchers Zhenke Duan and Xin Li have introduced THETA, a novel computational framework designed to solve the scalability problem in qualitative social science research. The system moves beyond simple frequency-based statistics by implementing Domain-Adaptive Fine-tuning (DAFT) using the LoRA technique on top of foundation embedding models. This process optimizes the semantic vector structures within specific social contexts—like financial regulation or public health—to capture latent meanings that traditional topic models often miss, a problem known as 'semantic thinning.'

To ensure the research maintains epistemological rigor, the team encapsulated the technical process within an AI Scientist Agent framework. This framework comprises three specialized agents: a Data Steward, a Modeling Analyst, and a Domain Expert. These agents work together to simulate the human-in-the-loop expert judgment and constant comparison processes that are central to established qualitative methods like grounded theory. Instead of producing a static output, the agents iteratively evaluate algorithmic clusters, perform cross-topic semantic alignment, and refine raw outputs into logically consistent theoretical categories.

Validation experiments across six distinct domains demonstrated that THETA significantly outperforms conventional models such as Latent Dirichlet Allocation (LDA), Embedded Topic Model (ETM), and Contextualized Topic Model (CTM) in both capturing domain-specific interpretive constructs and maintaining superior thematic coherence. By providing this capability as an open-source, interactive analysis platform, THETA aims to democratize advanced NLP for social scientists, ensuring the trustworthiness and reproducibility of findings derived from massive datasets like social media corpora, which were previously impractical to analyze manually.