Generalized word shift graphs: a method for visualizing and explaining pairwise comparisons between texts

Abstract A common task in computational text analyses is to quantify how two corpora differ according a measurement like word frequency, sentiment, or information content. However, collapsing the texts’ rich stories into single number often conceptually perilous, and it difficult confidently interpret interesting unexpected textual patterns without looming concerns about data artifacts validity. To better capture fine-grained differences between texts, we introduce generalized shift graphs, visualizations which yield meaningful interpretable summary of individual words contribute variation texts for any measure that can be formulated as weighted average. We show this framework naturally encompasses many most commonly used approaches comparing including relative frequencies, dictionary scores, entropy-based measures Kullback–Leibler Jensen–Shannon divergences. Through diverse set case studies ranging from presidential speeches tweets posted urban green spaces, demonstrate graphs flexibly applied across domains diagnostic investigation, hypothesis generation, substantive interpretation. By providing detailed lens shifts corpora, help social scientists, digital humanists, other analysis practitioners fashion more robust scientific narratives.