THETA: A Low-Threshold, High-Performance Topic Analysis Platform for Social Science Research

THETA is an open-source topic analysis platform specifically designed for social science research, integrating LLM semantic understanding capabilities with classic topic modeling methods. Its core values include low-threshold entry (beginners can start in 5 minutes), highly flexible configuration (hierarchical parameter control), and strong scientific standards (seven golden indicators). It supports three modes: zero-shot, fine-tuning, and supervision, integrates 12 baseline models for comparison, and provides an enterprise-level solution for social science text mining.

In social science research, text topic modeling is a core issue. Traditional models like LDA struggle to handle large-scale text. In recent years, LLM semantic embedding technology has brought new possibilities to topic modeling, and THETA emerged to address the need for social science researchers to conduct efficient analysis without a deep ML background.

The technical architecture is inclusive: the embedding layer uses Alibaba's Qwen series models (0.6B/4B/8B options, supporting Chinese and English) and also supports lightweight SBERT; the modeling layer implements a self-developed generative topic model while integrating 12 baseline models (including LDA, HDP, STM, ETM, BERTopic, etc.), covering traditional statistical, neural, and clustering methods.

Three modes are supported: Zero-shot mode (directly inferring with pre-trained Qwen, suitable for exploratory/small data); LoRA fine-tuning mode (lightweight parameter adjustment, balancing resources and performance); Supervised mode (training with labeled tags to improve accuracy).

It enforces seven golden standard indicators (topic diversity, iRBO, NPMI, C_V coherence, UMass coherence, exclusivity, perplexity), and multi-dimensional evaluation avoids single indicator bias. It supports multiple input data formats (txt/csv/docx/pdf), and the preprocessing pipeline automatically completes steps like cleaning and tokenization, reducing the workload of data preparation.

It generates rich visualizations (topic network diagrams, heatmaps, word clouds, radar charts, etc.) and supports Chinese-English switching. It provides a model selection decision tree: HDP/BERTopic is recommended when the number of topics is unknown; if known, choose based on text length, covariates, etc. Recommendations for models in different scenarios: choose LDA for speed, THETA+Qwen for quality, and multiple models for comparative research.

It supports local (conda+Python3.10+GPU) and cloud deployment, and provides installation scripts. Performance optimization suggestions: reduce batch size and model size. Academic citations are available in BibTeX format, community support is via email and GitHub, and it uses the Apache-2.0 license.