Summary
The episode works through four questions that students often find difficult in computational text analysis, cultural domain analysis and language-model-assisted coding. After opening with an invented language-model summary that contains a quotation no participant said, it uses an invented keyness comparison to show how raw counts mislead when transcripts differ in length and how one talkative participant can produce an apparently significant group keyword, and it notes the construction choices behind semantic network communities. Sarah calculates Smith's salience for five invented free lists and corrects a common error in the denominator, and the hosts use Statistics Canada loneliness data and an invented eigenvalue ratio to show that consensus analysis measures shared belief and can reveal two age-based cultural models. In the harder example, a reworded codebook definition introduced partway through coding nearly triples how often a code is applied and creates a spurious age difference, and the episode closes with a debate on how far language-model coding can be trusted, ending in a rule to delegate codes one at a time after validation, stability checks, audit and disclosure.