Definition
Topic modeling is unsupervised machine learning on text. Given a corpus, a topic model outputs a set of topics — each topic represented as a distribution over words — and a mapping from each document to the topics it contains. Classical methods include LDA (Latent Dirichlet Allocation), published by Blei, Ng, and Jordan in 2003, and NMF (Non-negative Matrix Factorization). Modern methods include BERTopic, which clusters sentence-transformer embeddings, and LLM-assisted clustering, which uses a large language model to group and label semantic clusters.
The important property is that no labels are required. The model looks at the corpus and decides what the topics are. That is exactly the property that makes topic modeling useful for feedback, and exactly the property that makes its output noisier than a tuned supervised classifier.
Why it matters
A consumer brand's feedback corpus changes every week. New defects surface, new accessories ship, new marketing campaigns reshape the language customers use, new retailers get added. A supervised classifier trained last quarter on last quarter's labels will miss a defect that emerged three weeks ago — and no brand has the labeling capacity to keep one fresh.
Topic modeling fills that gap. It runs on the current corpus, surfaces clusters that were not in any prior taxonomy, and gives analysts the raw material they need to refine via thematic analysis. In practice, production feedback systems combine topic modeling for discovery with zero-shot classification for routing and a stable taxonomy for reporting. Indellia's Theme Agent runs this pipeline so analysts see emerging topics the week they emerge, not the quarter after. Running a topic model on SKU-resolved feedback — not a brand-wide dump — also sharpens the clusters, because the linguistic signature of a coffee maker problem is different from a blender problem.
Example
A consumer-electronics brand runs a topic model on 38,000 reviews collected across Amazon, Best Buy, and Costco for a wireless earbud line. BERTopic surfaces 67 clusters. A Consumer Insights analyst inspects them: most map to known themes (battery, fit, sound), but cluster 41 is new — 340 reviews referencing pairing failures after a specific iOS update. The cluster did not exist in the brand's supervised taxonomy. The team escalates to firmware, promotes "iOS 18.1 pairing failure" into the active taxonomy, and notifies CX to update the response playbook — all within the same week the reviews appeared. A supervised-only system would have tagged the reviews under "pairing" generically, and the iOS-specific pattern would have surfaced only after a manual read of hundreds of tickets.