How have music emotions been described in Google books? Historical trends and corpus differences

The study investigates how emotions associated with music are described in written language, distinguishing perceived (expressed by music) and felt (evoked by music) emotions. Given music’s centrality to human life and its applications in recommendation, therapy, and information retrieval, the authors ask whether large text corpora can reveal people’s attitudes toward musical emotions over long historical periods and across cultures. They posit that language use in descriptions of music can reflect both real depictions and idealized expectations. Building on prior corpus-based cultural and psychological research, the paper poses three research questions: (1) What are the historical trends in the emotional description of music in books over the past two centuries? (2) Are these trends consistent across languages (American English, British English, Simplified Chinese)? (3) Do descriptions differ across corpora types (fiction vs overall books)?

Prior work has mined music-related texts to study meanings of artists, music journalism, and metaphors, but systematic analysis of emotional descriptions of music in texts is lacking. Corpus-based studies have traced historical shifts in personality descriptors, emotional expression, cultural values, morality, and women’s status, showing that language reflects social change and can be influenced by historical events. Cultural and linguistic research indicates cross-language differences in emotion expression, categorization, speech segmentation, and rhetorical patterns, motivating cross-linguistic comparisons. Even within a single language, corpora differ: adjectives vary across Twitter vs books; fiction uses more person-descriptive and archaic words and shows shifting abstract/concrete vocabularies; social media platforms differ in emotion norms. These lines suggest that both cultural differences and corpus composition may shape how music’s emotions are described.

Data source: Google Books Ngram (GBN) datasets v2. Historical trends were analyzed using the English corpus (~1,510,000 books), focusing on 1800–2000 (few books pre-1800; selection criteria changed post-2000). To measure importance of descriptors, the study used usage frequency of adjective–music bigrams (“adjective music”), following prior corpus methods. Adjective selection and polarity: 320 emotion-related English adjectives primarily from LIWC “affect” (with some from music emotion studies) were used; adjectives included both emotion descriptors (e.g., happy, sad) and emotion-associated words (e.g., sweet, terrible). Each adjective was categorized as positive, negative, or other per LIWC categories. Adjusted frequency: To control for varying base frequency of “music,” adjusted frequency for each adjective–music combination in year i was AF_combi = F_combi / F_music, where F_combi is the bigram frequency and F_music is frequency of the unigram “music.” Historical smoothing used a 7-year moving average (e.g., 1947–1953 for 1950). Category-level time series were computed by summing adjusted frequencies across adjectives within each polarity. Language comparison: American English (~1,160,000 books), British English (~370,000), and Simplified Chinese (~300,000) corpora were compared over 1960–2000 (Simplified Chinese appears in the 1950s). English analyses used the same 320 adjectives; Simplified Chinese analyses used 585 affect adjectives from the Simplified Chinese LIWC. Same adjusted-frequency methodology was applied. Corpus comparison: The Google English Fiction corpus (~330,000 fiction books) was compared with the overall English corpus using the same 320 adjectives and adjusted-frequency approach. Statistical analyses: Time series comparisons and tests included Wilcoxon signed-rank tests (reporting Z and effect size r), Friedman tests with post hoc Wilcoxon for multi-language comparisons, and Pearson correlations between overall adjective usage and adjective–music usage across years to assess whether trends were music-specific.

General trend specificity: Pearson correlations between overall adjective usage and its usage describing music over 1800–2000 were low on average (mean r = 0.172 ± 0.316), suggesting music-specific trends rather than mirroring general language usage. Historical trends in English (1800–2000): Positive adjectives were used more frequently than negative adjectives to describe music (Z = 12.293, p < 0.001, r = 1.000), consistent with a positivity bias. Usage of both positive and negative emotion-related adjectives declined over the two centuries despite “music” itself increasing in absolute and relative frequency. The most frequent descriptors included sweet, solemn, beautiful, fine, best, excellent, sad, and melancholy. Within positives, evaluative terms (fine, best, excellent) exceeded positive emotion words (happy, joyful, relaxing); among negatives, emotion words (sad, melancholy) were most frequent. The top word sweet strongly tracked the overall positive trend (r(200) = 0.945, p < 0.001), but removing sweet did not alter the overall trend direction. Negative adjectives such as sad, melancholy, strange, and bad showed slight declines after 1900; serious spiked in the early 19th century then declined later. Language differences (1960–2000): In AE, BE, and SC books, positive > negative adjectives when describing music (each: Z = 5.579, p < 0.001, r = 1.000). No overall difference across languages in positive adjective use (Friedman χ²(2) = 0.146, p = 0.929). Significant differences for negatives (Friedman χ²(2) = 58.390, p < 0.001): negative adjectives were less used in SC than AE and BE (post hoc Wilcoxon all p < 0.001, r = 1.000). SC showed a sharp surge in positive adjectives circa 1966–1976 (coinciding with China’s Cultural Revolution), during which SC exceeded AE (Z = 2.134, p = 0.033, r = 0.455); AE vs BE also differed (Z = 2.845, p = 0.004, r = 0.607), BE vs SC was not significant (p = 0.182). Trends over time within 1960–2000: Positive adjectives declined in BE (Z = 3.920, p < 0.001, r = 0.620) and modestly in AE (Z = 2.091, p < 0.05, r = 0.331); negative adjectives declined in AE (Z = 3.920, p < 0.001, r = 0.488) and BE (Z = 3.845, p < 0.001, r = 0.439) but not significantly in SC (Z = 0.597, p = 0.550). AE–BE adjective time series correlations were low on average (mean r = 0.141 ± 0.225) with some high word-specific correlations (e.g., sweet r = 0.839). Corpus differences (English fiction vs overall, 1800–2000): Positive > negative adjectives in fiction (Z = 13.394, p < 0.001, r = 0.940). Both positive and negative adjectives were more frequent in fiction than overall English books (positive: Z = 11.003, p < 0.001, r = 0.781; negative: Z = 10.862, p < 0.001, r = 0.771). Contextual interpretations: Peaks and declines aligned with historical periods (e.g., romantic era increase in early 19th century; later declines; serious elevated around eras including WWII, Great Depression, Baby Boom).

Findings consistently show a positivity bias in descriptions of music across languages and corpora, aligning with the Pollyanna hypothesis and the general prevalence of positive events. Despite this bias, the long-term decline in both positive and negative emotion-related adjectives suggests reduced emotional description of music in books over two centuries, potentially reflecting cultural shifts (e.g., collectivism to individualism) or changing expressive norms. Language differences indicate fewer negative descriptors in Simplified Chinese, consistent with East–West differences in emotional display rules and potential ideological or stylistic influences on music discourse. The spike in SC positives during 1966–1976 likely reflects sociopolitical influences of the Cultural Revolution. Corpus comparisons confirm that fiction is more emotion-laden than general books, in line with prior findings on fiction’s adjective use and intuition bias. While some trends correlate with general language usage, the low average correlations indicate music-specific dynamics. The results extend prior cultural and historical linguistics by focusing on music emotion descriptors, linking linguistic trends to cultural periods and events, and demonstrating cross-linguistic and corpus effects.

This study maps how emotion-related adjectives describing music have been used in Google Books across time, languages, and corpora. It reveals a robust positivity bias, long-term declines in emotional descriptors in English books (1800–2000), language differences with especially low negative usage in Simplified Chinese, and higher emotional adjective usage in fiction versus overall books. These patterns suggest that cultural shifts and historical events are reflected in music-related language. Future research should broaden targets beyond adjective–music bigrams (e.g., nouns, verbs, multiword expressions), include additional seed concepts (song, melody), extend to more languages and contemporary corpora, and leverage improved, representative datasets to clarify drivers of historical change and cultural differences.

The analysis is limited to adjective–music bigrams, excluding other grammatical forms and multiword constructions (e.g., predicates or intensifiers). The sole seed target “music” omits related concepts (song, melody), potentially narrowing scope. Corpora largely cover the last two centuries and may not reflect current usage. Reliance on LIWC lexicons, which are not exhaustive, may miss relevant vocabulary. Google Books Ngram poses reliability and representativeness concerns (OCR errors, lack of metadata, genre misclassification, only ~6% of published books, unstable corpus composition over time), complicating diachronic comparisons. Publication processes, censorship, and editorial choices can bias what appears in books; book language is more conservative than speech, so results reflect published rather than spoken trends. Simplified Chinese results pre-1978 are unstable due to limited book counts, warranting caution.