Topic modelling
Explore recurring word patterns across documents
Topic modelling estimates latent patterns of word co-occurrence and represents documents through combinations of topics. Latent Dirichlet Allocation models documents as mixtures over topics and topics as distributions over words. Topics are statistical structures that researchers interpret and label. Their meaning depends on the corpus, preprocessing, model assumptions and validation against documents or external information.
Use it for exploratory mapping of a sizeable corpus when recurring themes or document mixtures are relevant, and you can inspect source texts and compare plausible model specifications.
Strengths
- Summarises large corpora while allowing documents to mix topics
- Supports exploration of patterns across time or metadata
Limitations
- Topic labels are interpretations rather than discovered truths
- Topic number, stopwords and random initialisation can alter results
Know the boundary
Topic modelling is not equivalent to reflexive thematic analysis, and a topic label is not an observed variable.