Description
Giving each machine-derived topic a meaningful, defensible name by reading its top words and representative documents, ideally with multiple analysts and documented rationale.
Topic models output word lists and document weights, not meanings. Labelling is the interpretive step that turns them into research findings, and it is where most errors enter: labels based on top words alone can misrepresent what documents actually say. A rigorous protocol has at least two analysts read top words, distinctive words (e.g. FREX or c-TF-IDF) and 10-20 representative documents per topic, independently propose labels and short definitions, reconcile disagreements and mark topics that are incoherent or artefactual. Large language models can propose candidate labels, but analysts should verify them against documents. Labels should be descriptive rather than prematurely evaluative.
Attributes
- Discipline
- Qualitative Analysis & Synthesis
- Loss type
- Latent
- Stage
- Definition
- Data
- Qualitative
- Effort
- Hours
Steps
- For each topic, extract top words, distinctive words and the 10-20 highest-weighted documents.
- Have two analysts independently write a label (2-6 words) and a one-sentence definition from the documents, not only the words.
- Compare labels, discuss disagreements and record the agreed label with rationale.
- Mark topics as coherent, mixed or junk (e.g. greetings, transcription artefacts) and exclude junk from reporting.
- Optionally use an LLM to propose labels from the same evidence, and accept them only after analyst verification.
- Publish a topic codebook with labels, definitions, example quotes and prevalence.
Sources
- gensim (LDA)
- BERTopic
- sentence-transformers
- pyLDAvis (visualisation)
- hdbscan
- DiMaggio, P., Nag, M. & Blei, D. (2013) Exploiting affinities between topic modeling and the sociological perspective on culture. Poetics
- Lau, J. H. et al. (2011) Automatic labelling of topic models. ACL
- Saldaña, J. (2021) The Coding Manual for Qualitative Researchers, 4th ed.
- The Inquiry Series, C11 — Topic Modelling at Scale — LDA, BERTopic & Semantic Clustering
Also known as
- Topic interpretation
- Topic naming
Categories
- User Research
- Qualitative Analysis & Synthesis