Does typicality drive Cultural Success?

Grant Packard (Lazaridis) Jonah Berger (Wharton) Why do some cultural items succeed?

Two possibilities #1: Success is random

Even domain experts struggle predicting success1

Driven by unpredictable social drift2

1 Bielby & Bielby 1994; Hirsch 1972 2 Hahn & Bentley 2003; Salganik, Dodds & Watts 2006 #2: Individual level psychological processes shape collective outcomes

Psychology Culture à CulturePsychology2 1

1 Markus & Kitayama 1991 2 Kashima 2008; Norenzayan et al 2006; Schaller & Crandall 2004 Cultural Selection

Success depends on “fit” between item characteristics and shared human psychology

(Heath, Bell & Sternberg 2001; Kashima 2008; Schaller & Crandall 2004) Cultural consumption driver?

The stimulation of novelty1

1Flavell, Miller & Miller 2001; Zuckerman 1979 Cultural referents set the bounds

Novelty depends on experience1

Differentiation is bounded, not infinite2

1 Lena & Peterson 2008 2 Boyer 1994; Norenzayan et al. 2006; Zuckerman 2016 Our prediction

Cultural content that is more atypical, or differentiated from its category peers, will be more successful. : How to quantify cultural content’s “fit”?

Solution: natural language processing Data on thousands of songs

Billboard/Nielsen digital sales • Individual preferences • Every 3 months for 3 years

Seven major genres • Christian, Country, Dance, Pop, Rap, Rock and R&B

Ranked songs by genre (1-50) • 4,200 song rankings • 1,879 unique songs

Artist & radio airplay

Complete song lyrics Method

1) Determine lyrical themes across songs, genres

2) Calculate lyrical differentiation

3) Test relationship of differentiation and success 1. Determine song & genre themes Latent Dirichlet allocation1 1. 2. 3. 4. Set # Word co-occurrence Words probabil. Songs are a probabil. of topics w/in & across songs appear in topics distribution of topics

.52 Hello love .34 “…I’ve forgotten how it our .17 felt before the world felt .12 .43 fell at our feet. There’s (…) such a difference…” .05

.16 Formation highway. 14 “…Earned all this country .12 money, but they never air .10 .39 take the country out (….) me. I got hot sauce in…” .45

.23 from the Bottom money .16 “…Now I’m on the road, million .12 half a million for a show, ballin .09 .02 and we. Started from…” (…) .75 1 Blei 2012 1. Determine song & genre themes

Ten topic result (Selected for reliability, interpretability, parsimony)

“Anger and “Body Violence” Movement” “Dance Moves” “Family” “Fiery Love” bad, dead, hate, body, bounce, bop, dab, mash, american, boy, burn, feel, fire, kill, slay clap, jump, shake nae, twerk daddy, mamma, heart, love whoa

“Girls and Cars” “Positivity” “Spiritual” “Street Cred” “Uncertain Love” car, drive, girl, feel, like, mmm, believe, grace, ass, b*tch, dope, aint, cant, love, kiss, road oh, yeah lord, one, soul rich, street need, never

2. Calculate lyrical differentiation

Adapt Ireland and Pennebaker’s (2010) language style matching

Lyrical diffs,g = 1- Sk=1…n [(|topick,s – topick,g|)/(topick,s + topick,g + .0001)]

Genres with more extreme Genres with less extreme variation across topics variation across topics 3. Regress success on differentiation

Lyrical differentiation is positively linked to song success (B = 6.45, p = .001).

16% increase in atypicality associated with a one- position chart rank improvement (e.g., #2 to #1) This result is robust to…

Alternatives Specifications • Artist • Different approaches to • Promotion calculating differentiation (e.g. Jensen-Shannon divergence, • Time squared differences) • Similarity to other genres • Different specifications • Topics, style1, diversity2 (e.g. prop odds, log transforms, etc.) • Non-linearity • Other language – # of words, complexity, major psych constructs3, top 100 words

1 Ireland & Pennebaker 2010 2 Zhang & Grabchak, 2016 3 Pennebaker, Boyd, Jordan, & Blackburn, 2015 Can we isolate the impact of song lyrics versus other attributes? Isolating lyrical differentiation

Vary differentiation, fix other song attributes

Multi-genre songs (N = 410) “Started from the Bottom” should perform better in the genre in which they’re more Lyrical differentiation differentiated from peers Rap: 30.1% R&B: 28.3%

Result: Songs perform better where they are more differentiated (B = 34.64, p = .003). Lyrical differentiation appears to be driving success for highly successful songs.

But could less popular songs also be more differentiated? Assessing song selection (top 50)

1. Vary song success, fix artist

Album: Matched comparison of lyrical Drake differentiation of less popular Top 50 song songs by same artist on same “Started from the Bottom” album (N = 1,050 songs). Non Top 50 song “Too Much”

Result: Top 50 more differentiated from genre than less popular songs (B = .02, p < .001). Assessing song selection (top 50)

2. Account for upper bound in ranks (n = 50)

– Truncated Gaussian (B = 14.41, p < .001)

3. Apply a model that assesses selection

– Two-stage Heckman 1. Lyrical diff. predicts top 50 (B = 2.32, p < .001) 2. Top 50 status does not affect the relationship between lyrical differentiation and song success (Inverse Mills l = -106.93, p = .55) Robust relationship between lyrical differentiation and popularity

Might the relationship vary by genre?

Lyrics might matter less in Dance1 (about beat not words)

Difference might be less important in Pop2 (about mainstreaming) 1 Fabbri 1981 2 Frith 1986 Results by genre consistent What’s the nature of differentiation?

Different ways to be different: less/more typical

Result: Using more atypical topics linked to success (p = .001). Varies within genre (e.g. Rock should add Street Cred; p = .02). Contributions

1. Cultural items that are more differentiated from their peers are more popular

2. Psychological processes can shape the items that constitute culture

3. NLP can help link micro (psychological) and macro (cultural) processes. Additional questions

• Do other lyrical features drive success? – “You” in Pop and Dance – Exploring different ways of engaging “you”

• Do other song attributes play distinct roles? – The “music” may be key to cultural categorization (genre assignment) – Lyrics may be a relatively easy route to novelty – E.g., remixes that add new lyrics to old tunes Does typicality drive Cultural Success?

Grant Packard (Lazaridis) Jonah Berger (Wharton) ppendix

Model 1, Model 2, artist, Model 3, Model 4, top Simple song, time language 100 words effect controls controls (1) (2) (3) (4) The main result

Lyric al differentiation 6.45 ** 8.01 ** 8.38 *** 7.84 ** (1.99) (2.45) (2.48) (2.66) Times charted 0.83 *** 0.69 *** 0.75 *** (0.10) (0.11) (0.11) Multi -genre count 4.79 *** 4.58 *** 5.24 *** (0.80) (0.69) (0.77) Radio airplay 11.17 *** 11.10 *** 11.09 *** (0.55) (0.55) (0.56) LIWC Dictionaries Word count 0.00 0.00 (0.00) (0.00) Six letter 0.02 -0.02 (0.07) (0.07) Cognitive words -0.06 0.03 (0.07) (0.08) Affect words 0.00 0.05 (0.08) (0.08) Social words -0.07 -0.01 (0.05) (0.06) Perceptual words 0.02 0.18 (0.08) (0.12) Motivation words -0.02 -0.03 (0.06) (0.06) Temporal words -0.05 -0.06 (0.05) (0.06) Relativity words 0.00 -0.05 (0.05) (0.05) Swear words 0.15 0.06 (0.20) (0.22) Fixed effects Artist/song No Yes Yes Yes Topic No Yes Yes Yes Time No Yes Yes Yes Top 100 words No No No Yes

Intercept 23.34 *** 38.25 *** 39.95 *** 33.03 *** (0.70) (1.75) (2.48) (6.40) Adjusted R 2 0.023 Marginal R2 0.142 0.146 0.175 Conditional R2 0.344 0.347 0.367

*** p < .001, ** p < .01, * p < .05; Coefficients predict reverse-coded song ranking.