Links that speak: The global language network and its PNAS PLUS association with global fame Shahar Ronena, Bruno Gonçalvesb,c,d, Kevin Z. Hua, Alessandro Vespignanib, Steven Pinkere, and César A. Hidalgoa,1 aMacro Connections, Media Lab, Massachusetts Institute of Technology, Cambridge, MA 02139; bDepartment of Physics, Northeastern University, Boston, MA 02115; cAix-Marseille Université, CNRS, CPT, UMR 7332, 13288 Marseille, France; dUniversité de Toulon, CNRS, CPT, UMR 7332, 83957 La Garde, France; and eDepartment of Psychology, Harvard University, Cambridge, MA 02138 Edited by Kenneth W. Wachter, University of California, Berkeley, CA, and approved November 13, 2014 (received for review June 11, 2014) Languages vary enormously in global importance because of editions of Wikipedia, and posting short messages on Twitter historical, demographic, political, and technological forces. How- (“tweets”) in multiple languages. These coexpressions define ever, beyond simple measures of population and economic power, networks (Fig. 1) that—even though not representative of the there has been no rigorous quantitative way to define the global world’s general population—represent a coarse map of the influence of languages. Here we use the structure of the networks links connecting the elites that participate of these three im- connecting multilingual speakers and translated texts, as expressed portant global forums, as social connections often require a in book translations, multiple language editions of Wikipedia, and shared language. Twitter, to provide a concept of language importance that goes In this paper, we map the global language networks (GLNs) beyond simple economic or demographic measures. We find that expressed in three large records of linguistic expression, and use the structure of these three global language networks (GLNs) the structure of these networks to determine the degree to which is centered on English as a global hub and around a handful each language is global. First, we look at a collection of more of intermediate hub languages, which include Spanish, German, than 2.2 million book translations compiled by UNESCO’s Index French, Russian, Portuguese, and Chinese. We validate the mea- Translationum project. This dataset allows us to map the network ’ sure of a language s centrality in the three GLNs by showing that it of book translations, which are produced by individuals with a exhibits a strong correlation with two independent measures of high literary capacity (authors and professional translators) and SCIENCES the number of famous people born in the countries associated are shaped by market forces, such as the demand for books in APPLIED PHYSICAL with that language. These results suggest that the position of different languages. Each translation from one language to an- a language in the GLN contributes to the visibility of its speakers other forms a connection. Next, we map the network of linguistic and the global popularity of the cultural content they produce. coexpressions expressed by the community of digitally engaged knowledge specialists that edit Wikipedia. Here, two languages networks | languages | culture | digital humanities | fame are connected when users that edit an article in one Wikipedia language edition are significantly more likely to also edit an ar- f the thousands of languages that have ever been spoken, ticle in another language edition. Finally, we map the network only a handful have become influential enough to be con- COGNITIVE SCIENCES O of linguistic coexpressions expressed in Twitter. Here, two lan- PSYCHOLOGICAL AND sidered global languages. However, how do we measure the global guages are connected when users that tweet in a language are influence of a language? What are the implications of a world in also significantly more likely to tweet in another language. which only a handful of languages are globally influential? These three networks allow us to map the paths of direct In the past, researchers have used a variety of measures to and indirect communication between speakers from different determine the global influence of a language. Several studies languages. Our method formalizes the intuition that certain have relied on measures that proxy the global influence of a language using the population and wealth of its speakers (1–4). Significance While wealth and population approximate a language’s in- fluence, as the dissemination of a language has historically re- quired a strong power base (5), such measures fail to capture the People have long debated about the global influence of lan- global influence of a language: often the speakers of a language, guages. The speculations that fuel this debate, however, rely — and their wealth, are locally concentrated, making the language on measures of language importance such as income and population—that lack external validation as measures of a locally influential rather than globally influential. ’ An alternative method to measure the global influence of language s global influence. Here we introduce a metric of a language’s global influence based on its position in the net- a language is to focus on who speaks that language, and in work connecting languages that are co-spoken. We show that particular, on how connected the speakers of that language are. the connectivity of a language in this network, after control- In the words of linguist David Crystal, “Why a language becomes ling for the number of speakers of a language and their in- a global language has little to do with the number of people who ’ ” come, remains a strong predictor of a language s influence speak it. It is much more to do with who those speakers are. (5) when validated against two independent measures of the In the past, Latin was the pan-European language, not because it cultural content produced by a language’s speakers. was the mother tongue of most Europeans, but because it was the language of the Roman Empire and later the language of the Author contributions: S.R. and C.A.H. designed and performed research; S.R., B.G., and K.Z.H. Catholic Church, scholars, and educators (5). The use of Latin by retrieved data; S.R., K.Z.H., and C.A.H. analyzed data; B.G., A.V., and S.P. advised on analytic well-connected elites set it apart from other languages and helped methods; and S.R., S.P., and C.A.H. wrote the paper. Latin endure as a universal language for more than 1,000 years. The authors declare no conflict of interest. However, can we use these ideas to identify which modern This article is a PNAS Direct Submission. languages are globally influential? If global languages are those Freely available online through the PNAS open access option. connecting international elites, then we can identify the global Data deposition: Visit language.media.mit.edu for datasets and additional visualizations. languages associated with particular elites by mapping their 1To whom correspondence should be addressed. Email: [email protected]. networks of multilingual coexpressions. Examples of multilingual This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. coexpressions include book translations, edits to multiple language 1073/pnas.1410931111/-/DCSupplemental. www.pnas.org/cgi/doi/10.1073/pnas.1410931111 PNAS Early Edition | 1of7 Fig. 1. Visualizations of the three GLNs. The three GLNs contain all language connections that involve at least six users (Twitter and Wikipedia) or six translations and that are significant with P < 0.01. languages are disproportionately influential because they provide English and Spanish. An idea conceived by a Vietnamese speaker, direct and indirect paths of translation among most of the world’s however, might only reach a Mapudungun speaker in south-central other languages. For example, it is easy for an idea conceived by Chile through a circuitous path that connects bilingual speakers of a Spaniard to reach an Englishman through bilingual speakers of Vietnamese and English, English and Spanish, and Spanish and 2of7 | www.pnas.org/cgi/doi/10.1073/pnas.1410931111 Ronen et al. 6 PNAS PLUS 8 8 10 10 10 ABCeng deu spa fra eng eng itarus 6 6 nldjpn msa spa msa porzhopol swe 10 por 10 por spa 4 nor fil tur finhun nld fil nld 10 ukr cathbscesdan msa korronhebara ara jpn jpn fas fra ara fra bulslkell kor ita kor ita tha vie slv thaafr tur deu afr tur epo lit est cat ces rus tha catces rus deu glgsqi lat ell ell azeeuslavhin 4 glg zhonordanswe 4 glg dan zhoswe maltam islmkdben hbs hbsnor fil telafr kat 10 swa sqi 10 sqi bar ltz gle cymhyebel vielav slk pol swa lav pol marurd hat ronfin vieslk ron fin argvec javscoswaastkurbremongujkanociuzbkazyid ukr hun hat hun scn vls nepsinglafaochv mlt eus mlt eus ukr aka hatsom uigmri baktgk aze slvfasbulestheb aze slvfasheb div new hausottahnapyormlgkabchuudm arctuk san Users Twitter hin est bul 2 bugfrp gotsswcebsnasrdewegrn mltcrhquesndasmmdfangori tat bod tamcymmkd cymtammkdhin tpifijdzoglv tet aymfur hawfulsrnkhmsmocherohsmechmbuakaaxalavapankir urdlit lit 10 ilocsbrupwln amhmyaromoss gle isl Users Twitter urd chr corchywol laocos pan kat pan glekatisl tir pdcbisxhosagormipkpapibonavtonpuskalsah pli bel bel Wikipedia Editors mol sin mal hye sin hyemal lbemyv 2 lao telkan ben 2 lao kan ben crekom 10 tel pampagkin linzul iku div guj 10 div guj kon bam bod kas abk khmmya yid myakhmbod yid nya run chr syr chr ndo amhiku iku amh lug kik R2=63% R2=40% R2=66% 0 0 ori 0 ori 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 10 10 10 10 10 10 10 10 10 10 10 10 10 10 Book Translations (From) Book Translations (From) Wikipedia Editors 5 6 6 10 10 10 DEF 4 10 4 4 10 10 3 10 2 10 2 2 10 10 1 Users Twitter Common Users Twitter Common 10 Common Wikipedia Common Editors R2=23% R2=41% 0 59% 0 0 10 SCIENCES 0 2 4 6 10 0 2 4 6 10 0 5 10 10 10 10 10 10 10 10 10 10 Book Translations Book Translations Common Wikipedia Editors APPLIED PHYSICAL Fig.