Extract representative quotes for each STM topic

Takes the result of sframe_run_stm_topics() (not a raw stm model object) and, for each topic, pulls the top n_quotes documents by topic-document probability using stm::findThoughts(), then maps each one back to its original respondent row via the mapping sframe_run_stm_topics() stored on result$fit.

Usage

extract_quotes(model, text, n_quotes = 3L)

Arguments

model

A stm_topics result list from sframe_run_stm_topics() (i.e. result, not result$fit and not the raw stm object).

text

The response vector the topic model was fit on: either the raw vector (one entry per original data row, indexed 1:1 by row number) or the clean_text_responses()-cleaned vector, which drops blank/missing rows and is therefore shorter, its positions no longer equal original row numbers once any earlier row was dropped. Both forms work correctly: when text carries the respondent attribute clean_text_responses() sets, that mapping is used to find each quote's real position; otherwise text is assumed to be the raw, 1:1-indexed vector.

n_quotes

Integer. Quotes to return per topic. Default 3L.

Value

A data.frame with columns topic, rank, respondent (the original row index in the data the model's text argument came from, not a document-matrix or corpus row index), and quote.