Topic modeling faces a fundamental granularity dilemma, according to a new arXiv preprint. Probabilistic topic models assign topics at the word level, while transformer-based methods operate at the document level. The abstract argues that the latter overlooks something about real-world documents, though the sentence is cut off in the available excerpt.
The paper's title indicates the proposed solution: sentence-level topic modeling that draws on both lexical and contextual evidence. This positions the work as an attempt to bridge the gap between fine-grained word-level models and coarse document-level representations.
Because the abstract is truncated, the full method and evaluation details are not available from the source. The contribution, as stated, is the identification of the granularity dilemma and a sentence-level approach that combines lexical and contextual signals.