Unsupervised = no labels needed. You feed in 10 million raw documents and out come K topics, each defined by its top words. Topic 1: {stock, market, trading, fund} → you say 'Finance.' Topic 2: {patient, surgery, diagnosis, hospital} → you say 'Healthcare.' The algorithm finds patterns; humans name them. This is the key trade-off: zero annotation cost, but topics require human interpretation and may not align with your desired categories.

This is one of LDA's well-known failure modes: 'background topics' or 'junk topics' that capture frequent, non-discriminative words. Words like 'said,' 'year,' 'would,' 'new' appear everywhere and don't belong to any specific topic. Solutions: (1) Remove more stopwords including domain-specific ones, (2) Use a smaller K, (3) Increase the number of Gibbs sampling iterations, (4) Use asymmetric priors. In practice, ~20% of LDA topics are often junk — interpreting only the good ones is part of the workflow.

Short text (3 sentences) is LDA's weakness — there aren't enough words per document to build reliable word co-occurrence statistics. NMF on TF-IDF is better but still surface-level. BERTopic shines here: (1) Sentence-BERT produces meaningful embeddings even for short text, (2) HDBSCAN finds clusters of varying density (some issues are common, others rare), (3) c-TF-IDF labels each cluster with discriminative terms. For 50K documents, BERTopic runs in minutes on a single GPU.

This is LDA's most common failure mode. Function words (the, is, of) appear in every document with roughly uniform frequency. LDA allocates a topic to capture this distribution, wasting a slot. Fix: (1) Remove standard stop words. (2) Remove words appearing in >90% of documents (corpus-specific stop words). (3) Remove words in <5 documents (too rare). (4) Use TF-IDF to filter low-information words. Some implementations like MALLET have built-in stop word optimization, but proper preprocessing is always essential.