The neuron fires most strongly on content words that are unusually informative or distinctive in their context (i.e. high–TF-IDF or “keyword”‐type tokens).