Glossary

Topic Modeling

An unsupervised machine learning method that discovers recurring themes in a text corpus — useful for exploring large collections of survey responses, interview transcripts, or documents.


Topic modeling is an unsupervised machine learning approach that discovers latent themes (topics) in a collection of documents without requiring pre-labeled training data. Given a corpus of texts, the model identifies groups of words that tend to appear together and represents each document as a mixture of these topics.

How it works:

Classic topic models (Latent Dirichlet Allocation, LDA) treat each document as a mixture of topics and each topic as a distribution over words. A document about climate policy might be 60% “policy/regulation” topic and 40% “emissions/science” topic, with neither label provided by a human — the model discovers the structure from word co-occurrence patterns.

Modern approaches like BERTopic use sentence transformers (contextual embeddings) to represent document meaning before clustering, producing more coherent and interpretable topics than LDA.

Research applications:

  • Analyzing open-ended survey responses at scale
  • Exploring themes in interview transcript collections
  • Mapping research trends across large paper corpora (arXiv, PubMed)
  • Policy document analysis
  • Social media corpus exploration before deeper qualitative work

Critical limitations for social science:

  • Topics are statistical, not semantic. A topic is a word cluster, not a human concept. Two very different ideas can cluster together because they share vocabulary; two closely related ideas can separate because they use different terminology.
  • Human interpretation is required. Topic models don’t label topics — a researcher must examine the top words and representative documents for each cluster and decide what the “topic” means. This is an interpretive act, not an automatic output.
  • Results vary with parameters. The number of topics (k) must be set by the researcher, and different choices produce different results. There is no objectively correct number of topics.
  • Not suitable as a standalone finding. Reporting “BERTopic identified 8 themes” without deeper qualitative engagement with those themes is methodologically insufficient for most social science venues.

Related terms: Structured Data Extraction

Related guide: Social Science