Answers are grounded in this paper and cite their sources.
Document clustering can make search results easier to explore, but phrase and semantic representations have different failure modes. Phrase n-grams produce inspectable labels yet can miss paraphrases, whereas latent or transformer representations can connect vocabulary variants but may be harder to explain. This study develops a reproducible hybrid clustering prototype that fuses phrase TF-IDF, a semantic representation, and lexical TF-IDF. The package includes K-means and agglomerative baselines, three channel ablations, a fusion-weight sweep, five controlled synthetic conditions, machine-readable assignments, metadata, visualisations, and tests. The bundled corpus contains 400 synthetic documents across four topical domains, with generated paraphrase and noise variation. In the primary run, several methods attain near-perfect partition scores; this is interpreted as a ceiling effect caused by the deliberately structured synthetic corpus, not evidence of broad superiority. The default semantic channel uses TF-IDF with truncated singular value decomposition (SVD); an optional Sentence Transformers mode is evaluated separately. The phrase component uses contiguous n-grams and is not a faithful compressed suffix-tree STC implementation. The contribution is a transparent, reproducible benchmark scaffold for evaluating complementarity, sensitivity, and failure modes. Validation on public real-world corpora, multiple independent seeds, transformer-specific runs, and human retrieval tasks remains necessary before drawing general conclusions.
No takes yet. Share an insight, caveat, or question.
Chile-Agada et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: