LDA (probabilistic, mixed-membership)
document
"stocks rose after
the fed cut rates"
topic mixture
{finance: 0.7,
politics: 0.2, ...}
topic_word matrix:
finance -> { stocks: 0.04, rate: 0.03, ... }
politics -> { fed: 0.05, vote: 0.04, ... }
tech -> { chip: 0.04, ai: 0.03, ... }
each row sums to 1
good for long docs, topic mixtures, limited compute.
weak on short text. no semantics.
interpretable probability distributions.
BERTopic (neural, cluster-based)
docs (N)
BERT embed
384-dim
UMAP
5-dim
HDBSCAN
cluster 0
1
2
3
class-based TF-IDF
top words per cluster = topic representation
one topic per document. short text friendly. semantic.