What it does

Top2Vec trains on a pile of text, figures out how many topics are actually in there, and builds a shared vector space where documents, words, and topics all live together. Once trained, you can query topics by keyword, find similar documents, or rummage through related words without building a separate search pipeline.

The interesting bit

The newer “Contextual Top2Vec” (beta) goes document-level down to token-level: it assigns topics to individual tokens and finds topic spans inside single documents. That means a long document covering, say, both cryptocurrency and sourdough can be segmented by theme rather than dumped into one bucket. The catch: it only works with two specific sentence-transformer models right now.

Key highlights

• Automatically detects topic count; no hand-tuning k like LDA
• No stopword lists or stemming required
• Joint embedding means topics, documents, and words are directly comparable in the same space
• Supports Doc2Vec, Universal Sentence Encoder, or BERT sentence transformers
• New contextual mode adds per-token topic assignments and document topic distributions
• Built-in semantic search: query documents by topic or keyword

Caveats

• Contextual Top2Vec is explicitly marked beta; the README warns of “issues or unexpected behavior”
• Contextual mode is locked to all-MiniLM-L6-v2 or all-mpnet-base-v2; no other embedding models
• Doc2Vec may outperform pretrained encoders on large or highly specialized vocabularies, but you’ll wait longer

Verdict

Worth a look if you need topic modeling without the usual hyperparameter grief, especially if semantic search over those topics is part of the job. Skip it if you need production-grade token-level segmentation today, or if your stack can’t accommodate the contextual mode’s limited model choices.