How To Calculate Language Skill Premia

Transcription

1 Language barriers to foreign trade: evidence from translation costs Alejandro Molnar Vanderbilt University DRAFT, NOT FOR CIRCULATION November 9, 2013 Abstract [Foreign trade involves tasks that may be subject to language barriers, such as researching foreign markets, communicating with counterparties and marketing products to foreign consumers. Language skills have a wage premium that is determined by local and worldwide supply and demand for language services, and the premium is specific to each country and pair of languages. I construct a novel measure of language skill premia based on professional rates for translation services from an online market. The skill premium measure relies on the bi-directional nature of the translation cost data to control both for difficulties inherent in defining a unit of account (as the per word piece-rates common in the translation industry do not embody equal amounts of work across languages) and the skilled-wage component of rates. I develop an estimation strategy based on overlaps in ethnolinguisitic populations to estimate the effect of the language skill premium as a cost barrier to trade, net of confounders such as trade by shared ethnic populations. I find that accounting for country and languagespecific language barriers yields a three-fold increase in the estimated effect of language on foreign trade, relative to current estimates based on a shared common language.] JEL: F1, F14, F23, Z13 Keywords: Gravity equation, common language, language barriers. I am grateful to Tim Bresnahan, Jon Levin and Kalina Manova for guidance and encouragement. Special thanks to Ran Abramitzky, Kyle Bagwell, Emilio Depetris Chauvin, Bernardo Díaz de Astarloa, Doireann Fitzgerald, Gordon Hanson, Han Hong, Paul Ma and Andrés Rodríguez-Clare for helpful comments and discussions. All remaining errors are my own. Department of Economics, Vanderbilt University, 415 Calhoun, Nashville, TN 37204, [email protected]. 1

2 1 Introduction In this paper I study the relationship between the languages that are spoken in a country and the country s patterns of foreign trade. Foreign trade involves tasks that may be subject to language barriers, such as researching foreign markets, communicating with counterparties and marketing products to foreign consumers. Language skills have a wage premium that is determined by local and worldwide supply and demand for language services, and the premium is specific to each country and pair of languages. The main idea in this paper is to exploit variation in country-specific prices for translation services across different pairs of languages to recover country-specific measures of language skill premia. I construct a novel measure of these premia based on professional rates for translation services from an online market. The skill premium measure relies on the bi-directional nature of the translation cost data to control both for difficulties inherent in defining a unit of account (as the per word piece-rates common in the translation industry do not embody equal amounts of work across languages) and the skilled-wage component of rates. I develop an estimation strategy based on overlaps in ethnolinguisitic populations to estimate the effect of the language skill premium as a cost barrier to trade, net of confounders such as trade by shared ethnic populations. I find that accounting for country and language-specific language barriers yields a threefold increase in the estimated effect of language on foreign trade, relative to current estimates based on a shared common language. 2 Online markets for translation services In this section I describe the offline and online markets for translation services, as well as the price data available from online translation markets and how it contains information on skill premia for a specific form of human capital that is used intensively in foreign trade. The task of translation is typically carried out by a single individual. A translator s physical productivity can be thought of as the pace at which a source text is translated into a target text of a given quality. A translator s productivity is text-dependent, as translation work may require domain-specific knowledge in addition to language skills, and text may vary in difficulty, requiring additional time (e.g. for research) to produce a translation of a given quality. The typical translator is a freelance worker, but relationships with the demand side of the translation market may be in-house as well as arms length. Relationships may 2

3 be long- or short-term, and occur directly or through intermediaries called translation agencies. Agencies provide demand risk-sharing among teams of individual translators, as well as reputation services (on reputation intermediaries in online labor markets, see Stanton and Thomas, 2011). The price of translation work is called a translation rate and is quoted in a unit of text that can be easily counted on a computer: words in spacing languages (i.e. those in which words are separated by white space such as English) and characters in non-spacing languages (e.g. Simplified Chinese). Translation rates are specific to a directed language pair (e.g. Spanish to English) and are almost always expressed in units of the source text, as quoting per unit in the target language provides bad incentives for the translator. As freelancers, translators negotiate rates bilaterally with potential clients and will usually quote client-specific rates based on the current state of demand for their services and the attributes of the client: for example, translators may quote higher rates for rushed jobs or highly technical jobs, or lower rates when providing quantity discounts or attempting to establish a relationship with a client that may be a future source of demand. Offline demand for a translator can come from professional listing services, translation agencies, reputation and word of mouth. A further source of demand can be outsourcing from other translators, and these translators provide editing and monitoring and may or may not disclose to the final client that the translation was outsourced. Most demand from these offline sources is specific to a translator s country of residence. Online platforms specifically designed to intermediate global translation markets started in 1999 (proz.com) and 2002 (translatorscafe.com). These platforms introduced new forms of market organization (e.g. procurement auctions, explicit reputation metrics for both sides of the market) and greatly facilitated the meeting of supply and demand across borders. To do business on these platforms, translators must create a profile and report the minimum translation rate at which they are willing to work on each language pair in which they work. These rates are not revealed to potential clients, but screen the jobs that a translator will see on a job listings dashboard when logged into the site. Since this screening rate is set prior to and independently of the attributes of listed translation jobs, it can be thought of as a reservation wage (in units of work, rather than time) for each translator. The second largest of these online markets, translatorscafe.com, discloses the average of this minimum translation rate for all translators in a language pair that are located in a particular country. There is substantial variation in these average rates both across language pairs within country and across countries within language pair. The determinants supply and 3

4 demand for translation services that give rise to variation across language pairs within country arises are reasonably straightforward: for example, demand may depend on the languages used by trade partners, foreign tourists, and on the country s interest in cultural products produced in foreign languages, and these in turn may depend on the country s ethnolinguistic composition. Supply follows from each country s endowment of people with language skills that enable work in each specific language pair (which may be relatively fixed in the short term), and an opportunity cost of time for this type of work, common across all possible translation pairs (e.g. the wage for the bundle of skills that translators possess net of their language-specific skills). Variation across countries in the same language pair is less straightforward, as a law of one price might be expected to hold in online markets. One reason for cross country variation in prices is that translation services are not homogeneous and may require country-specific knowledge (e.g. on the legal environment, pop culture, slang or vernacular) for which translators in different countries are imperfect substitutes. A second reason follows from the microstructure of the translation industry and the fact that translators have limited capacity and spend a fraction of their time unemployed and waiting for the arrival of the next job. Accepting a low-paying job removes the translator from the market until the job is completed, and therefore may preclude accepting a job that arrives later with a higher pay. Online translators face such arrival processes for potential jobs from both online and offline sources, and a simple search model suggests that they should set a reservation wage for accepting an online job that depends on the opportunity cost of removing themselves from the offline market for a period of time. A further idiosyncratic factor of this industry is that the unit of work in which prices are expressed is not constant across languages. Different western languages may differ in their use of articles, contractions and compound words, so that texts that are supposed to convey the same meaning would be counted at different lengths depending on the language in which they are written. For instance, the English phrase it is not so simple" consists of 5 words, whereas the same phrase in Spanish no es tan sencillo" consists of 4. 1 This applies even more clearly to rate comparisons between spacing and non-spacing languages, where rates are not expressed in words. 1 The phrase in English can be contracted to it s not so simple" or it isn t so simple", both of which are also 4 words long. On average, text in Spanish tends to be longer than equivalent text in English. German uses long compound words, and a famous example that arose from a state legislature was the Rindeischetikettierungsueberwachungsaufgabenuebertragungsgesetz", or the Law on delegation of duties for supervision of beef labeling". If the amount of work required from a translation is not a function of the number of words but of the amount of meaning conveyed in the text, we should expect, all else equal, that translation rates out of German be higher than out of English, and these higher than out of Spanish. 4

5 To develop a comparable measure across languages, I assume the following structure for observed translation rates, which are quoted in US dollar cents per source word or character: r abc = δ a δ (a,b)c η abc (1) where r abc is the observed average translation rate from source language a to target language b for translators located in country c, δ a is a scaling factor specific to the source language intended to absorb differences between languages in the amount of work embodied in translating a word or character, and δ (a,b)c is an undirected pair and country effect. I normalize log δ (eng,spa)usa = 0 and regress log r abc on source language fixed effects and undirected pair and country fixed effects, so average rates by source language relative to the rate on this specific pair are absorbed by source effects, and remaining variation in rates is captured by the δ (a,b)c term. For example, the value of δ (eng,spa)arg is constructed from the average of rates for English to Spanish (net of the English source language effect) and Spanish to English (net of the Spanish source language effect) for translators located in Argentina. After adjusting for source effects, a country s average translation rate on a language pair remains a nominal quantity. Equal nominal rates can represent widely different resource costs for countries with different average wage levels for skilled labor. example, the average English to French rate is 10 USD cents per word for translators located in France, 13 cents in Côte d Ivoire, 12 cents in Algeria, 10 cents in Morocco, 8 cents in Cameroon and 7 cents in Senegal. As France s GDP per capita is approximately 40 times that of Senegal, the resource cost of employing a person to overcome English to French language barriers is presumably substantially higher as a per-person share of Senegal s economy than it is for the French economy, and much more so for the economy of Cote d Ivoire. 2 To obtain a measure of variation across language pairs that is net of wages, I regress log δ (a,b)c on GDP per capita in country c, and take the residuals of this regression as my final measure of real language skill premia. Figure?? plots this exercise, plotting log δ (a,b)c in blue for language pairs that involve a country s most widely spoken language, and green otherwise. Some of the highest language skill premia include English-French in Cote d Ivoire, English-Swahili in Tanzania, and Italian-Japanese in 2 Similarly, translation rates between Scandinavian languages are relatively high, presumably because most speakers of these languages reside in high wage countries. A reasonable prior is that language barriers between Scandinavian counties are relatively low. For 5

6 Japan. Some of the lowest include English-Farsi in Afghanistan, English-Khmer in Cambodia, Kazakh-Russian in Kazakhstan and English-Albanian in Greece. Figure 2 plots a subset of this data, narrowing in on countries for which English is not the most widely spoken language and the skill premium for English and the country s most widely spoken language. 2.1 Why language could matter in foreign trade Trade requires communication, which can give rise to language barriers. Languageintensive tasks that are essential to trade include researching foreign markets, adapting and marketing products to foreign consumers, 3 and communication and contracting between importing and exporting firms. Language may affect the foreign direct investment decisions of multinational firms, e.g. the location of regional headquarters, and trade in final and intermediate goods may follow from such decisions. Empirical estimates of trade costs acknowledge the role of language (?), usually estimated by inclusion of a binary variable for whether two countries share an official language. The size of language barriers can be expected to depend on the languageintensity of specific tasks involved in trade, and the cost of hiring workers within a country with the required language skills. 3 Empirical evidence on language barriers to trade In this section I describe how my measure of language skill premia explains trade flows between countries by including the language skill premia described in Section 2 in standard trade gravity estimation frameworks from the trade literature. I also describe an empirical strategy to estimate a causal effect of language barriers on trade. In order to include the language and country-specific adjusted translation rates in a standard gravity equation framework, I map the adjusted rates to country pairs in two ways. I define the top pair country-pair specific rate as the average of the rates between the most widely spoken languages in each country. I also assign a value of 1 to 3 Firms make design or product choices in response to language barriers, an example of which are the text-free assembly manuals that accompany furniture sold across many national markets by Swedish firm IKEA. To sidestep individual language costs, IKEA incurs the cost of high quality design in assembly manuals (and perhaps product adaptation) to avoid ambiguity and the use of written text. Not all retail products that require instruction manuals or assembly instructions in a language other than that of design or manufacture are suitable to, or have sufficient scale to afford, IKEA s economy of words. See Kelly and Zetzsche (2012) for examples of how translation services are employed in trade. 6

7 the translation rate observed dummy for the pair. 4 I construct a fractional or populationweighted adjusted translation rate by applying the above procedure to every language population within a country combined with every language population in a partner country. The translation rate measure is obtained from the weighted average of every cell with a combination of languages, where the weight is the product of the marginal population measures in each country. Since rates for most potential language pairs are not observed, the measure of population for which the rate is observed is counted in the continuous translation rate observed variable. 5 Additional linguistic variables include the measure of the population in a country pair that speak the same language, i.e. the probability that a person picked at random from one country would be able to speak with a person picked at random from the other. This is equivalent to the main measure used in the preceding work on the effect of language on trade by Melitz (2008) and Melitz and Toubal (2012). I include the log of this variable and a dummy for when countries share no speakers of a common language. I also include the dummy variable for common official language, which is the usual control for language used in almost all prior empirical work that estimates the gravity equation. Table 1 reports OLS estimates for three specifications of the standard gravity equations for the subset of country pairs with positive trade flows. The first column includes only the standard common language dummy, the second includes the fractional adjusted rate measures and the third only the rates for the top pair of languages. 6 From the estimates in Column (2), a 1 percent increase in the adjusted translation rate between the languages of a pair of countries is associated with a 0.6 percent decline in bilateral trade. Column (3) presents a weaker, non-significant estimate of the same effect for the most widely spoken language for each country in the pair. In both cases the magnitude of estimates for distance, contiguity and colonial relationships all decline substantially after inclusion of the larger set of language skill premium and linguistic overlap measures. The magnitude for common official language is almost halved, but inclusion of the measures of ethnolinguistic overlap means that this binary variable identifies purely the official status of any common languages. Table 2 presents results from an exponential regression on the same data, which allows inclusion of pairs with zero trade that are dropped due to the log transformation, and is a favored empirical method in the trade literature because gravity equation 4 If language pair data for only one country is observed, I include the rate and count it as observed. 5 Some translation rates in the data that are plotted in Figure?? do not map to any ethnic population (e.g. English to Spanish in Norway) and are therefore not used in gravity equation estimates. 6 Results are robust to netting GDP per capita from nominal rates with a quadratic term or a local polynomial regression. 7

8 regressions are motivated by multiplicative structural models and estimates from the exponential regression framework are robust to heteroscedasticity in the error term for the parameters of interest in these models. Both this framework and the framework I will employ below to instrument for the language measures cannot include importer and exporter fixed effects, so I include instead a remoteness measure (used for example by Baldwin and Harrigan (2011) and Manova and Zhang (2012), see discussion in Head and Mayer (2013)). The main coefficient estimates on this sample are similar in magnitude, but the estimates from the Column (3) specification are now significant. Common official language becomes non-significant, and the inclusion of the full set of language variables has a smaller effect on the coefficients for geographical covariates such as distance and contiguity. The language barrier estimates from Tables 1 and 2 cannot be given the causal interpretation of a trade cost because linguistic overlap can be correlated with trade through other channels, in particular ethnic trade (Rauch and Trindade, 2002). If trade creates additional demand for translation services on the relevant language pairs and this increases translation rates, this would lead trade to be positively correlated with observed translation rates, biasing upwards (in this case, towards attenuation) the coefficient estimates for translation rates as a trade cost in a gravity equation. I use data on the overlaps of ethnolinguistic populations to develop a shifter for the relative scarcity of language skills and estimate a causal effect of country-specific language skills on trade. To describe the empirical strategy, consider the example of trade between the United Kingdom and either Vietnam or Thailand. The population of Vietnamese and Thai speakers in the UK is small and similar in magnitude, but there is one large difference between these language pairs from the perspective of the UK market for language services: the existence of the US as a majority English-speaking country with a large population of Vietnamese speakers. There is no similar example for the Thai language. The existence of the US as a country where Vietnamese and English speakers overlap has an effect on the world market for language services in this language pair. In particular, if the ethnolinguistic composition of the US has a larger effect on supply than on demand for English-Vietnamese language services, this is likely to decrease the English-Vietnamese language skill premium in the UK. Based on this idea, I define an ethnic overlap instrument in the following manner: for a given language in country" pair (e.g. English in the UK, Vietnamese in Vietnam), the probability that a speaker of each of these languages would coexist in an country other than that of the language pair with a speaker of the other language. That is, I take the measure of all English speakers worldwide with the exception of the UK, and 8

9 calculate the conditional probability that if they were to meet a fellow resident at random from their country, that resident would be a speaker of Vietnamese. This conditional probability is close to zero for most English speakers worldwide, except for the US where it is about 0.02, and the US constitutes a large fraction of worldwide English-speakers. I then do the same for Vietnam, where the measure is almost zero in neighboring countries with Vietnamese speaking residents, but a large fraction of Vietnamese speakers outside the UK-Vietnam pair reside in the US, where their probability of meeting an English speaker is close to 1. probabilities. I construct the instrument as the product of both of these As this instrument coarsens the variation in the data to a language-pair level (since foreign overlap will be a similar value for most countries, e.g. the instrument has a similar value for English-Vietnamese pairs that involve Vietnam and Great Britain, Australia or New Zealand), the estimated effect might be expected to be similar to that from an OLS regression of trade flows on a language pair (but not country) specific measure. However, whereas the average global rate may be driven by trade flows on a particular country (e.g. the abundance of Vietnamese speakers in the US affects both the worldwide supply and demand for English-Vietnamese language services, as well as trade between the US and Vietnam through a direct ethnic channel, see Rauch and Trindade, 2002), the instrument specifically excludes the correlation between trade flows and a country pair s common language fraction (e.g., in the case of US to Vietnam trade, the instrument takes on a low value, whereas the common language fraction takes on a relatively high value). The fact that I do not observe language skill premia in all language pairs presents a hurdle to straightforward instrumental variables estimation, as the common practices of selecting a sample or interacting unobserved or censored values of a regressor with a dummy variable (as I do in the results presented in Tables 1 and 2) are only valid under exogeneity assumptions. When a censored regressor is endogenous, 2SLS estimation can lead to a bias that amplifies the magnitudes of estimated coefficients, as discussed by Rigobon and Stoker (2007). 7 As detailed in Appendix Appendix B, I implement the estimation strategy of Chernozhukov, Rigobon, and Stoker (2010) to estimate the effect of language skill premia on trade using the ethnolinguistic overlap instrument. Results are in Table 3, both for the conditional on positive linear regression specification and the exponential regression. The bottom row in the table presents results for the first stage: for the language 7 In fact, estimating Columns 2 and 3 of Table 1 by 2SLS with the ethnolinguistic overlap instrument does lead to implausibly large effects of language skill premia on trade flows 9

10 pair of a specific country, ethnolinguistic overlap elsewhere in the world predicts a lower median adjusted translation rate, which suggest that the effects of ethnolinguistic overlap through the expansion of supply dominates the effects through expansion of demand. Any effect from the ethnolinguistic overlap for the common pair in question is absorbed by the common language fraction variable, which is included in the second stage. The main coefficient estimates for the effect of language skill premia on trade are larger in magnitude than in the specifications of Tables 1 and 2, which is consistent with the main concern for endogeneity in those regressions being attenuation from the reverse the effect of trade on translation costs. The magnitudes from the exponential regression specification imply that predicted trade for a pair of countries sharing a common language is 5.2 times that of a pair of countries not sharing a common language and having the largest language skill premium observed in the data, relative to a 1.7- times increase estimated using only the common language dummy. This is a threefold increase in the estimated effect of language on foreign trade. Results are robust across specification that use alternative thresholds for translator data quality and when I include data from a second source (translationdirectory.com, an online listing directory for translators, where incentives to price revelation are not as clear). [To be done: Classification of products (e.g. more differentiated products, contractintense, R&D intense, etc.). Preliminary results show monotonic results between product differentiation or technological component and impact of measure. Pairs through English as a lingua franca. Extensive vs. intensive margins with WB Exporter Dynamics data. Alternative instruments based on migration flows and linguistic cleavages (e.g. Shastry, 2012). List of potential alternative channels, e.g. Warcziarg-style covariates, and Guiso, Sapienza, and Zingales (2009). Recast specification into the ideal for comparability of OLS and control function results, which is to splinter the gravity equation into population-proportional subcells to use the population-weighted language barrier measure between country pairs. For OLS show equivalence of splinter-regression to standard language-population-weighted] 4 Conclusions This paper develops a novel measure of language barriers between countries based on prices for translation services, which reflect the market premium on scarce language skills, and estimates the impact of this measure on trade flows between countries. The paper s main result is that the conventional practice of controlling for language with a common official language dummy omits a large share of the effect of language 10

11 on foreign trade. Understanding the proper role of language contributes to our understanding of the barriers to trade and economic integration that some countries may face: much as landlocked countries trade less, firms in countries with low endowments of foreign language skills may face additional hurdles to carrying out the multiple activities involved in foreign trade, and its domestic firms may need to rely on the initiative of foreign partners to overcome these barriers. Additional knowledge on the component factors that are regularly proxied by distance in standard gravity applications may reduce the relevance attributed to this catch-all variable, and increase our understanding of the nuanced factors that affect foreign trade. As an institutional and cultural endowment, the abundance of language skills may have broader implications for other flows such as the transmission of technological knowledge and cultural values. 11

12 References Baldwin, R., and J. Harrigan (2011): Zeros, Quality, and Space: Trade Theory and Trade Evidence, American Economic Journal: Microeconomics, 3(2), Chernozhukov, V., and H. Hong (2002): Three-Step Censored Quantile Regression and Extramarital Affairs, Journal of the American Statistical Association, 97(459), Chernozhukov, V., R. Rigobon, and T. M. Stoker (2010): Set identification and sensitivity analysis with Tobin regressors, Quantitative Economics, 1(2), Guiso, L., P. Sapienza, and L. Zingales (2009): Cultural Biases in Economic Exchange?, The Quarterly Journal of Economics, 124(3), Hallak, J. C. (2006): Product quality and the direction of trade, Journal of International Economics, 68(1), Head, K., and T. Mayer (2013): Gravity Equations: Workhorse, Toolkit, and Cookbook, CEPR Discussion Paper No Kelly, N., and J. Zetzsche (2012): Found in Translation. Perigee Books. Kugler, M., and E. Verhoogen (2012): Prices, Plant Size, and Product Quality, The Review of Economic Studies, 79(1), Manova, K., and Z. Zhang (2012): Export Prices Across Firms and Destinations, The Quarterly Journal of Economics, 127(1), Melitz, J. (2008): Language and foreign trade, European Economic Review, 52(4), Melitz, J., and F. Toubal (2012): Native language, spoken language, translation and trade,. Rauch, J. E. (1999): Networks versus markets in international trade, Journal of International Economics, 48(1), Rauch, J. E., and V. Trindade (2002): Economics and Statistics, 84(1), Ethnic Chinese networks in international trade, Review of Rigobon, R., and T. M. Stoker (2007): Estimation with Censored Regressors: Basic Issues*, International Economic Review, 48(4), Shastry, G. K. (2012): Human Capital Response to Globalization Education and Information Technology in India, Journal of Human Resources, 47(2), Silva, J. S., and S. Tenreyro (2006): The log of gravity, The Review of Economics and Statistics, 88(4), Stanton, C., and C. Thomas (2011): Landing the First Job: The Value of Intermediaries in Online Hiring, SSRN Scholarly Paper ID , Social Science Research Network, Rochester, NY. 12

13 Tables and figures [To do: table of main languages, language iso codes, number of translators, average rate against English, Spanish, French, Chinese, English in the US.] Figure 1: Adjusted translation cost and per capita GDP amh eng ETH eng tir ETH eng fra MDG eng swa TZA eng fas AFG eng pus AFG eng fra RWA eng nep NPL eng fra HTI eng hat HTI ben eng BGD eng swa KEN eng som KEN fra swa KEN eng fra KEN ara eng KEN deu eng KEN eng khm KHM eng tgk TJK eng rus TJK ara eng PSE eng rus KGZ eng fra SEN fra spa SEN eng pus PAK eng urd PAK ara urd PAK eng fra CIV eng lao LAO ara eng YEM deu vie VNM eng vie VNM vie zho VNM eng zho VNM fra vie VNM ara eng SDN hin tam IND eng fra IND hin mal IND dan eng IND eng mal IND ara eng IND hin jpn IND eng san IND hin kan IND eng kor IND eng kan IND deu eng IND fra hin IND hin pan IND guj hin IND bih eng IND eng guj IND eng nep IND deu hin IND hin mar IND ben hin IND hin tel IND eng zho IND eng jpn IND ara urd IND eng mar IND eng pan IND hin urd IND eng urd IND eng por IND eng nld IND eng spa IND eng ita IND eng hin IND hin rus IND eng tam IND eng fas IND ben eng IND eng rus IND eng tel IND eng hau NGA eng yor NGA rus uzb UZB eng uzb UZB eng rus UZB eng tgk UZB eng spa NIC ron rus MDA fra rus MDA eng rus MDA eng ron MDA deu rus MDA fra ron MDA eng spa HND eng nld PHL eng spa PHL deu eng PHL eng zho PHL eng zlm PHL dan eng PHL eng jpn PHL eng tgl PHL eng ilo PHL eng fra PHL deu spa BOL eng spa BOL ara deu EGY eng ita EGY ara fra EGY ara tur EGY ara heb EGY ara rus EGY ara eng EGY ara spa EGY ara ita EGY eng tur EGY ara zho EGY eng sin LKA eng tam LKA ara eng SYR deu fra MAR ara deu MAR fra spa MAR eng fra MAR ara spa MAR ara eng MAR fra ita MAR ara fra MAR eng rus MNG mon rus MNG eng mon MNG eng spa GTM eng rus GEO kat rus GEO eng kat GEO eng hye ARM hye rus ARM deu rus ARM eng rus ARM eng ind IDN eng jav IDN eng zlm IDN eng zho IDN deu ind IDN fra ind IDN ind zho IDN ind jpn IDN eng kur IRQ ara eng IRQ deu ukr UKR fra ukr UKR deu eng UKR eng ukr UKR spa ukr UKR ell rus UKR rus spa UKR rus zho UKR ita ukr UKR eng rus UKR fra rus UKR ara rus UKR pol ukr UKR pol rus UKR deu rus UKR ces rus UKR ita rus UKR heb rus UKR jpn rus UKR rus ukr UKR eng spa PRY eng spa SLV eng sqi ALB ita sqi ALB eng fra TUN ara fra TUN ara eng TUN fra spa ECU eng spa ECU ara spa JOR ara eng JOR ara fra JOR eng hbs BIH eng hbs MKD fra mkd MKD eng sqi MKD deu mkd MKD ita mkd MKD eng mkd MKD eng zho THA eng tha THA fra tha THA jpn tha THA deu tha THA ara eng DZA eng fra DZA ara fra DZA eng spa CUB fra spa CUB eng rus CHN ita zho CHN eng fra CHN eng zho CHN spa zho CHN eng ita CHN eng nld CHN jpn zho CHN eng jpn CHN eng spa CHN rus zho CHN fra zho CHN kor zho CHN deu zho CHN ara eng CHN eng kor CHN deu eng CHN ara eng LBY deu eng DOM eng fra DOM fra spa DOM eng spa DOM ita rus BLR deu rus BLR bel eng BLR bel deu BLR rus spa BLR bel rus BLR eng rus BLR fra rus BLR pol rus BLR ita spa PER por spa PER deu spa PER eng spa PER fra spa PER eng hbs SRB hbs ita SRB eng sqi SRB deu hbs SRB eng hun SRB rus srp SRB eng mkd SRB eng rus SRB fas fra IRN eng kur IRN deu fas IRN eng fas IRN aze eng AZE eng rus AZE eng tur AZE aze rus AZE por spa COL ita spa COL fra spa COL eng spa COL eng hbs MNE eng rus BGR bul fra BGR deu rus BGR bul ell BGR bul eng BGR bul spa BGR bul ita BGR bul rus BGR bul deu BGR eng spa PAN deu eng ZAF eng nld ZAF eng fra ZAF eng spa ZAF eng ita ZAF eng zul ZAF eng por ZAF afr eng ZAF eng rus ZAF eng nld SUR fra spa CRI eng spa CRI eng fra MUS hun ron ROU fra ron ROU ron spa ROU eng rus ROU deu ron ROU deu hun ROU ron rus ROU eng ron ROU eng hun ROU fra ita ROU por ron ROU ita ron ROU deu eng ROU fra hun ROU eng ita ROU eng fra LBN ara eng LBN ara spa LBN ara fra LBN eng ind MYS eng jpn MYS eng tam MYS eng fra MYS ara eng MYS eng zlm MYS eng zho MYS zho zlm MYS jpn zlm MYS ind zlm MYS por spa MEX eng spa MEX eng rus MEX deu spa MEX eng swe MEX fra spa MEX eng fra MEX eng ita MEX rus spa MEX spa swe MEX ita spa MEX deu eng MEX eng fas TUR rus tur TUR deu tur TUR ita tur TUR aze eng TUR fra tur TUR spa tur TUR ara tur TUR ara eng TUR eng tur TUR eng kur TUR eng spa VEN ita spa VEN por spa VEN deu spa VEN fra spa VEN spa zho ARG eng spa ARG deu fra ARG deu spa ARG por spa ARG eng ita ARG ita spa ARG deu eng ARG rus spa ARG fra spa ARG eng kaz KAZ kaz rus KAZ eng rus KAZ deu eng BRA por spa BRA eng ita BRA eng por BRA deu por BRA ita por BRA fra por BRA lav rus LVA eng rus LVA deu lav LVA eng lav LVA fra lav LVA nld rus RUS rus zho RUS fra rus RUS fin rus RUS pol rus RUS eng ukr RUS deu eng RUS rus srp RUS jpn rus RUS rus swe RUS rus tur RUS eng rus RUS rus spa RUS ara rus RUS ces rus RUS rus ukr RUS nor rus RUS eng zho RUS deu rus RUS por rus RUS ita rus RUS fra lit LTU eng rus LTU eng lit LTU lit pol LTU deu lit LTU ita lit LTU lit rus LTU eng rus POL pol spa POL nld pol POL ces pol POL dan pol POL pol por POL eng pol POL pol rus POL pol ukr POL deu pol POL deu eng POL eng ukr POL fra pol POL ita pol POL deu eng URY eng por URY por spa URY ita spa URY fra spa URY eng spa URY deu spa URY deu eng HUN eng ron HUN hun ron HUN hun nld HUN deu hun HUN hun ita HUN eng hun HUN fra hun HUN hun spa HUN hun rus HUN hrv ita HRV eng ita HRV eng hbs HRV eng slv HRV deu hrv HRV deu eng HRV fra spa CHL por spa CHL eng spa CHL deu spa CHL ita spa CHL jpn spa CHL eng est EST est rus EST eng rus EST deu est EST eng spa TTO eng slk SVK fra slk SVK ita slk SVK ces slk SVK ces eng SVK deu slk SVK hun slk SVK ces deu SVK slk spa SVK eng hun SVK rus slk SVK eng zho TWN ara eng SAU ara fra SAU deu eng CZE ces pol CZE ces slk CZE ces eng CZE eng rus CZE deu slk CZE ces ita CZE ces spa CZE ces fra CZE ces rus CZE ces deu CZE ces nld CZE eng slk CZE eng ita MLT eng fra MLT eng mlt MLT ita mlt MLT deu eng MLT eng por PRT deu por PRT nld por PRT por spa PRT por rus PRT ita por PRT fra por PRT por ron PRT eng jpn KOR eng zho KOR jpn kor KOR eng kor KOR ita slv SVN slv spa SVN deu slv SVN fra slv SVN deu eng SVN hrv slv SVN eng slv SVN ell rus GRC eng fra GRC ell sqi GRC eng spa GRC deu eng GRC ell nld GRC eng rus GRC bul eng GRC eng srp GRC ell ita GRC deu ell GRC bul ell GRC eng sqi GRC ell spa GRC ell eng GRC eng ita GRC ell fra GRC fra spa PRI eng spa PRI eng tur CYP ell eng CYP ell fra CYP fra heb ISR eng nld ISR ara heb ISR eng heb ISR heb spa ISR eng spa ISR heb rus ISR deu eng ISR eng ita ISR eng ron ISR eng rus ISR fra rus ISR deu heb ISR eng fra ISR ara eng ISR ara spa ESP eng spa ESP spa ukr ESP fra spa ESP cat fra ESP cat rus ESP por spa ESP jpn spa ESP ell spa ESP bul spa ESP nor spa ESP slk spa ESP eng glg ESP rus spa ESP nld spa ESP cat spa ESP pol spa ESP deu spa ESP eus spa ESP cat deu ESP spa swe ESP ces spa ESP ron spa ESP dan spa ESP fra glg ESP spa tur ESP ita spa ESP spa zho ESP cat eng ESP glg spa ESP cat ita ESP eng eus ESP eng ron ITA eng jpn ITA deu fra ITA ces ita ITA fin ita ITA eng zho ITA eng hun ITA bul eng ITA cat ita ITA ita swe ITA ita srp ITA ita jpn ITA dan ita ITA ita tur ITA ell ita ITA hbs ita ITA dan eng ITA ita zho ITA eng fra ITA ita rus ITA fra nld ITA ita ukr ITA eng slv ITA ita sqi ITA eng sqi ITA fra ron ITA eng por ITA eng swe ITA eng lat ITA fra por ITA eng pol ITA ita nor ITA ita pol ITA eng nld ITA ita lit ITA deu rus ITA eng srp ITA eng hbs ITA eng rus ITA fra rus ITA ita por ITA hrv ita ITA ara eng ITA deu ita ITA fra ita ITA fra spa ITA eng ukr ITA deu eng ITA ita ron ITA eng spa ITA ita nld ITA eng hrv ITA ita slv ITA hun ita ITA ita spa ITA ara fra ITA bul ita ITA ara ita ITA eng ita ITA eng fin ITA ita slk ITA deu eng NZL eng zho NZL eng jpn NZL eng rus NZL eng spa NZL eng fra NZL eng nld NZL eng ukr GBR bel eng GBR eng slk GBR eng pol GBR eng heb GBR eng nld GBR eng lav GBR fra por GBR eng fas GBR eng ron GBR ita rus GBR eng hbs GBR eng lit GBR fra ita GBR por spa GBR eng hrv GBR eng kur GBR eng lat GBR eng zho GBR ita spa GBR ben eng GBR eng pan GBR eng vie GBR eng guj GBR ara fra GBR eng tha GBR eng fra GBR afr eng GBR ita pol GBR eng ita GBR eng kat GBR dan eng GBR eng ind GBR cat eng GBR eng urd GBR eng hun GBR deu ita GBR eng glg GBR eng fin GBR eng tur GBR deu ell GBR eng nor GBR eng pus GBR ita por GBR eng sqi GBR eng prs GBR eng est GBR bos eng GBR eng srp GBR ces eng GBR ara eng GBR eng spa GBR eng jpn GBR eng slv GBR eng uzb GBR eng rus GBR eng kor GBR bul eng GBR ell fra GBR deu eng GBR eng por GBR lav rus GBR eng hin GBR eng swe GBR eng zlm GBR ell eng GBR ell spa GBR fra swe FRA ces fra FRA fra por FRA eng nld FRA fas fra FRA fra hun FRA por spa FRA fra tur FRA fra ron FRA fin fra FRA eng fra FRA deu ita FRA bul fra FRA eng por FRA fra spa FRA nld spa FRA ita spa FRA ell fra FRA fra srp FRA fra jpn FRA cat fra FRA deu nld FRA fra nor FRA fra nld FRA fra rus FRA fra ukr FRA eng tur FRA ara fra FRA fra slk FRA deu fra FRA eng ita FRA fra zho FRA fra pol FRA dan fra FRA fra ita FRA ara eng FRA eng isl ISL deu por DEU deu swe DEU ces deu DEU rus spa DEU eng fra DEU fra spa DEU deu nor DEU eng pol DEU eng jpn DEU eng tur DEU deu fra DEU eng por DEU eng lit DEU deu rus DEU deu srp DEU por spa DEU eng ind DEU ara deu DEU bul deu DEU deu hun DEU deu spa DEU ces eng DEU fra rus DEU eng swe DEU eng heb DEU eng spa DEU deu pol DEU cat deu DEU eng ita DEU deu eng DEU eng nld DEU ara eng DEU eng fin DEU deu vie DEU ell eng DEU eng rus DEU fra por DEU eng ukr DEU deu ron DEU eng ron DEU eng fas DEU deu jpn DEU deu tur DEU deu zho DEU cat spa DEU ita spa DEU deu fin DEU deu hrv DEU deu nld DEU fra nld DEU deu ell DEU eng zho DEU bul eng DEU deu ukr DEU dan deu DEU cat eng DEU fra ita DEU eng vie DEU deu slk DEU eng hun DEU deu ita DEU ara fra ARE ara eng ARE eng hin ARE eng rus ARE eng fas ARE eng fra ARE eng ita JPN eng zho JPN fra jpn JPN deu jpn JPN deu eng JPN ita jpn JPN eng spa JPN jpn zho JPN eng fra JPN jpn spa JPN eng ind JPN eng rus JPN eng jpn JPN ind jpn JPN eng ind SGP eng jpn SGP eng zlm SGP eng zho SGP eng spa SGP ara eng SGP eng fra SGP eng vie SGP nld por BEL deu nld BEL eng fra BEL fra ron BEL fra por BEL nld swe BEL fra ita BEL eng ita BEL nld spa BEL ara eng BEL eng nld BEL ell fra BEL fra rus BEL ita nld BEL fra spa BEL ita spa BEL eng spa BEL fra nld BEL deu fra BEL deu eng BEL nld rus BEL ara fra BEL deu rus BEL eng por BEL ben eng USA eng kor USA eng hrv USA pol rus USA bul eng USA fra rus USA ell fra USA amh eng USA eng mya USA ell eng USA eng ron USA eng nld USA eng mar USA eng fin USA eng pus USA eng rus USA eng ukr USA eng vie USA por spa USA cat eng USA eng ind USA deu rus USA fra tur USA eng kur USA bel eng USA eng hin USA eng hye USA eng lit USA eng tel USA eng prs USA eng tha USA afr eng USA deu fra USA eng ita USA fra nld USA eng som USA bos eng USA ara eng USA eng pol USA eng zlm USA jpn kor USA ita spa USA eng pan USA eng tgk USA deu ron USA dan eng USA eng mkd USA deu ita USA eng slv USA eng spa USA eng sqi USA eng srp USA fra ron USA deu eng USA eng slk USA eng est USA eng swe USA cat spa USA eng swa USA bul rus USA deu spa USA jpn zho USA eng hbs USA eng lav USA eng hat USA eng uzb USA rus spa USA eng khm USA eng tam USA eng fas USA eng guj USA fra spa USA eng fra USA heb rus USA eng heb USA aze eng USA eng por USA eng zho USA ita rus USA fra zho USA eng kat USA eng yid USA fra por USA eng mal USA ces eng USA ara fra USA eng tgl USA jpn spa USA eng hun USA eng nor USA fra ita USA rus ukr USA eng jpn USA eng mon USA eng urd USA eng lat USA eng tur USA eng lao USA eng nep USA fra hat USA eng gle IRL eng ita IRL eng slk IRL eng rus IRL eng ron IRL eng pol IRL ces eng IRL deu eng IRL ara eng IRL eng por IRL eng nld IRL eng fra IRL eng lit IRL eng spa IRL eng swe IRL eng nor IRL eng zho IRL eng ita FIN deu eng FIN fin rus FIN eng swe FIN fin spa FIN eng spa FIN fin fra FIN eng rus FIN fin ita FIN fin swe FIN eng fin FIN deu fin FIN deu spa AUT deu fra AUT deu rus AUT deu nld AUT deu eng AUT deu hrv AUT ces deu AUT eng fra AUT deu ita AUT deu hun AUT deu jpn AUT nld swe NLD nld spa NLD deu nld NLD eng tur NLD ita nld NLD nld pol NLD fra nld NLD nld tur NLD hun nld NLD nld por NLD eng nld NLD nld rus NLD fra jpn CAN eng fas CAN eng pan CAN eng lat CAN fra ita CAN eng hbs CAN eng ron CAN deu fra CAN fra spa CAN ita spa CAN eng ita CAN eng swe CAN eng hun CAN eng nld CAN fra por CAN fra zho CAN fra ron CAN deu eng CAN eng nor CAN ces eng CAN ell eng CAN fra rus CAN eng zho CAN eng kur CAN eng swa CAN eng ind CAN eng rus CAN eng jpn CAN eng spa CAN dan eng CAN eng tur CAN eng vie CAN ara fra CAN eng kor CAN eng slk CAN eng hrv CAN eng srp CAN eng pus CAN eng heb CAN eng urd CAN eng pol CAN eng fra CAN eng hin CAN eng por CAN bul eng CAN eng ukr CAN deu spa CAN eng prs CAN ara eng CAN dan eng SWE nld swe SWE spa swe SWE eng fin SWE fin swe SWE deu swe SWE eng swe SWE rus swe SWE eng spa SWE jpn swe SWE dan swe SWE fra swe SWE ita swe SWE nor swe SWE dan eng DNK eng fra DNK dan pol DNK eng spa DNK eng rus DNK eng swe DNK dan swe DNK dan deu DNK eng pol DNK eng nor DNK dan fra DNK dan spa DNK dan ita DNK deu eng DNK dan rus DNK dan nor DNK eng zho AUS deu eng AUS eng vie AUS eng por AUS eng ita AUS ara eng AUS eng spa AUS eng pol AUS eng fra AUS eng nld AUS ita spa AUS eng rus AUS eng ind AUS eng kor AUS eng jpn AUS ara eng KWT eng por CHE fra spa CHE eng rus CHE ara fra CHE eng spa CHE deu ita CHE eng swe CHE eng zho CHE eng ita CHE deu spa CHE deu eng CHE fra rus CHE ara eng CHE eng fra CHE fra ita CHE deu fra CHE ara eng QAT dan eng NOR eng swe NOR eng spa NOR eng rus NOR nor spa NOR nor swe NOR dan nor NOR deu fra LUX fra spa LUX eng fra LUX deu eng LUX amh eng ETH eng fas AFG eng nep NPL eng hat HTI ben eng BGD eng khm KHM eng tgk TJK ara eng PSE eng lao LAO ara eng YEM deu vie VNM eng vie VNM vie zho VNM fra vie VNM ara eng SDN hin tam IND hin mal IND hin jpn IND hin kan IND fra hin IND hin pan IND guj hin IND deu hin IND hin mar IND ben hin IND hin tel IND hin urd IND eng hin IND hin rus IND rus uzb UZB eng uzb UZB eng spa NIC ron rus MDA eng ron MDA fra ron MDA eng spa HND eng tgl PHL deu spa BOL eng spa BOL ara deu EGY ara fra EGY ara tur EGY ara heb EGY ara rus EGY ara eng EGY ara spa EGY ara ita EGY ara zho EGY eng sin LKA ara eng SYR ara deu MAR ara spa MAR ara eng MAR ara fra MAR mon rus MNG eng mon MNG eng spa GTM kat rus GEO eng kat GEO eng hye ARM hye rus ARM eng jav IDN ara eng IRQ deu ukr UKR fra ukr UKR eng ukr UKR spa ukr UKR ita ukr UKR pol ukr UKR rus ukr UKR eng spa SLV eng sqi ALB ita sqi ALB ara fra TUN ara eng TUN fra spa ECU eng spa ECU ara spa JOR ara eng JOR ara fra JOR eng hbs BIH fra mkd MKD deu mkd MKD ita mkd MKD eng mkd MKD ara eng DZA ara fra DZA eng spa CUB fra spa CUB ita zho CHN eng zho CHN spa zho CHN jpn zho CHN rus zho CHN fra zho CHN kor zho CHN deu zho CHN ara eng LBY fra spa DOM eng spa DOM bel eng BLR bel deu BLR bel rus BLR ita spa PER por spa PER deu spa PER eng spa PER fra spa PER eng hbs SRB hbs ita SRB deu hbs SRB fas fra IRN deu fas IRN eng fas IRN aze eng AZE aze rus AZE por spa COL ita spa COL fra spa COL eng spa COL eng hbs MNE bul fra BGR bul ell BGR bul eng BGR bul spa BGR bul ita BGR bul rus BGR bul deu BGR eng spa PAN eng zul ZAF eng nld SUR fra spa CRI eng spa CRI hun ron ROU fra ron ROU ron spa ROU deu ron ROU ron rus ROU eng ron ROU por ron ROU ita ron ROU ara eng LBN ara spa LBN ara fra LBN eng zlm MYS zho zlm MYS jpn zlm MYS ind zlm MYS por spa MEX eng spa MEX deu spa MEX fra spa MEX rus spa MEX spa swe MEX ita spa MEX rus tur TUR deu tur TUR ita tur TUR fra tur TUR spa tur TUR ara tur TUR eng tur TUR eng spa VEN ita spa VEN por spa VEN deu spa VEN fra spa VEN spa zho ARG eng spa ARG deu spa ARG por spa ARG ita spa ARG rus spa ARG fra spa ARG kaz rus KAZ eng rus KAZ por spa BRA eng por BRA deu por BRA ita por BRA fra por BRA lav rus LVA deu lav LVA eng lav LVA fra lav LVA nld rus RUS rus zho RUS fra rus RUS fin rus RUS pol rus RUS rus srp RUS jpn rus RUS rus swe RUS rus tur RUS eng rus RUS rus spa RUS ara rus RUS ces rus RUS rus ukr RUS nor rus RUS deu rus RUS por rus RUS ita rus RUS fra lit LTU eng lit LTU lit pol LTU deu lit LTU ita lit LTU lit rus LTU pol spa POL nld pol POL ces pol POL dan pol POL pol por POL eng pol POL pol rus POL pol ukr POL deu pol POL fra pol POL ita pol POL por spa URY ita spa URY fra spa URY eng spa URY deu spa URY hun ron HUN hun nld HUN deu hun HUN hun ita HUN eng hun HUN fra hun HUN hun spa HUN hun rus HUN eng hbs HRV fra spa CHL por spa CHL eng spa CHL deu spa CHL ita spa CHL jpn spa CHL eng est EST est rus EST deu est EST eng slk SVK fra slk SVK ita slk SVK ces slk SVK deu slk SVK hun slk SVK slk spa SVK rus slk SVK eng zho TWN ara eng SAU ara fra SAU ces pol CZE ces slk CZE ces eng CZE ces ita CZE ces spa CZE ces fra CZE ces rus CZE ces deu CZE ces nld CZE eng mlt MLT ita mlt MLT eng por PRT deu por PRT nld por PRT por spa PRT por rus PRT ita por PRT fra por PRT por ron PRT jpn kor KOR eng kor KOR ita slv SVN slv spa SVN deu slv SVN fra slv SVN hrv slv SVN eng slv SVN ell rus GRC ell sqi GRC ell nld GRC ell ita GRC deu ell GRC bul ell GRC ell spa GRC ell eng GRC ell fra GRC fra spa PRI eng spa PRI ell eng CYP ell fra CYP fra heb ISR ara heb ISR eng heb ISR heb spa ISR heb rus ISR deu heb ISR ara spa ESP eng spa ESP spa ukr ESP fra spa ESP por spa ESP jpn spa ESP ell spa ESP bul spa ESP nor spa ESP slk spa ESP rus spa ESP nld spa ESP cat spa ESP pol spa ESP deu spa ESP eus spa ESP spa swe ESP ces spa ESP ron spa ESP dan spa ESP spa tur ESP ita spa ESP spa zho ESP glg spa ESP ces ita ITA fin ita ITA cat ita ITA ita swe ITA ita srp ITA ita jpn ITA dan ita ITA ita tur ITA ell ita ITA hbs ita ITA ita zho ITA ita rus ITA ita ukr ITA ita sqi ITA ita nor ITA ita pol ITA ita lit ITA ita por ITA hrv ita ITA deu ita ITA fra ita ITA ita ron ITA ita nld ITA ita slv ITA hun ita ITA ita spa ITA bul ita ITA ara ita ITA eng ita ITA ita slk ITA deu eng NZL eng zho NZL eng jpn NZL eng rus NZL eng spa NZL eng fra NZL eng nld NZL eng ukr GBR bel eng GBR eng slk GBR eng pol GBR eng heb GBR eng nld GBR eng lav GBR eng fas GBR eng ron GBR eng hbs GBR eng lit GBR eng hrv GBR eng kur GBR eng lat GBR eng zho GBR ben eng GBR eng pan GBR eng vie GBR eng guj GBR eng tha GBR eng fra GBR afr eng GBR eng ita GBR eng kat GBR dan eng GBR eng ind GBR cat eng GBR eng urd GBR eng hun GBR eng glg GBR eng fin GBR eng tur GBR eng nor GBR eng pus GBR eng sqi GBR eng prs GBR eng est GBR bos eng GBR eng srp GBR ces eng GBR ara eng GBR eng spa GBR eng jpn GBR eng slv GBR eng uzb GBR eng rus GBR eng kor GBR bul eng GBR deu eng GBR eng por GBR eng hin GBR eng swe GBR eng zlm GBR ell eng GBR fra swe FRA ces fra FRA fra por FRA fas fra FRA fra hun FRA fra tur FRA fra ron FRA fin fra FRA eng fra FRA bul fra FRA fra spa FRA ell fra FRA fra srp FRA fra jpn FRA cat fra FRA fra nor FRA fra nld FRA fra rus FRA fra ukr FRA ara fra FRA fra slk FRA deu fra FRA fra zho FRA fra pol FRA dan fra FRA fra ita FRA eng isl ISL deu por DEU deu swe DEU ces deu DEU deu nor DEU deu fra DEU deu rus DEU deu srp DEU ara deu DEU bul deu DEU deu hun DEU deu spa DEU deu pol DEU cat deu DEU deu eng DEU deu vie DEU deu ron DEU deu jpn DEU deu tur DEU deu zho DEU deu fin DEU deu hrv DEU deu nld DEU deu ell DEU deu ukr DEU dan deu DEU deu slk DEU deu ita DEU ara fra ARE ara eng ARE fra jpn JPN deu jpn JPN ita jpn JPN jpn zho JPN jpn spa JPN eng jpn JPN ind jpn JPN eng zho SGP ben eng USA eng kor USA eng hrv USA bul eng USA amh eng USA eng mya USA ell eng USA eng ron USA eng nld USA eng mar USA eng fin USA eng pus USA eng rus USA eng ukr USA eng vie USA cat eng USA eng ind USA eng kur USA bel eng USA eng hin USA eng hye USA eng lit USA eng tel USA eng prs USA eng tha USA afr eng USA eng ita USA eng som USA bos eng USA ara eng USA eng pol USA eng zlm USA eng pan USA eng tgk USA dan eng USA eng mkd USA eng slv USA eng spa USA eng sqi USA eng srp USA deu eng USA eng slk USA eng est USA eng swe USA eng swa USA eng hbs USA eng lav USA eng hat USA eng uzb USA eng khm USA eng tam USA eng fas USA eng guj USA eng fra USA eng heb USA aze eng USA eng por USA eng zho USA eng kat USA eng yid USA eng mal USA ces eng USA eng tgl USA eng hun USA eng nor USA eng jpn USA eng mon USA eng urd USA eng lat USA eng tur USA eng lao USA eng nep USA eng gle IRL eng ita IRL eng slk IRL eng rus IRL eng ron IRL eng pol IRL ces eng IRL deu eng IRL ara eng IRL eng por IRL eng nld IRL eng fra IRL eng lit IRL eng spa IRL eng swe IRL eng nor IRL eng zho IRL fin rus FIN fin spa FIN fin fra FIN fin ita FIN fin swe FIN eng fin FIN deu fin FIN deu spa AUT deu fra AUT deu rus AUT deu nld AUT deu eng AUT deu hrv AUT ces deu AUT deu ita AUT deu hun AUT deu jpn AUT nld swe NLD nld spa NLD deu nld NLD ita nld NLD nld pol NLD fra nld NLD nld tur NLD hun nld NLD nld por NLD eng nld NLD nld rus NLD eng fas CAN eng pan CAN eng lat CAN eng hbs CAN eng ron CAN eng ita CAN eng swe CAN eng hun CAN eng nld CAN deu eng CAN eng nor CAN ces eng CAN ell eng CAN eng zho CAN eng kur CAN eng swa CAN eng ind CAN eng rus CAN eng jpn CAN eng spa CAN dan eng CAN eng tur CAN eng vie CAN eng kor CAN eng slk CAN eng hrv CAN eng srp CAN eng pus CAN eng heb CAN eng urd CAN eng pol CAN eng fra CAN eng hin CAN eng por CAN bul eng CAN eng ukr CAN eng prs CAN ara eng CAN nld swe SWE spa swe SWE fin swe SWE deu swe SWE eng swe SWE rus swe SWE jpn swe SWE dan swe SWE fra swe SWE ita swe SWE nor swe SWE dan eng DNK dan pol DNK dan swe DNK dan deu DNK dan fra DNK dan spa DNK dan ita DNK dan rus DNK dan nor DNK eng zho AUS deu eng AUS eng vie AUS eng por AUS eng ita AUS ara eng AUS eng spa AUS eng pol AUS eng fra AUS eng nld AUS eng rus AUS eng ind AUS eng kor AUS eng jpn AUS ara eng KWT ara eng QAT log avg. undirected pair translation rate net of source f.e log country GDP per capita Vertical axis: log of the average translation rates between a pair of languages (in both directions), net of source language fixed effects). English-Spanish rates for translators located in the United States are normalized to zero. Sample is coded by color: language pairs that involve a country s most widely spoken language in blue, language pairs that do not involve a country s most widely spoken language in green. Black line is linear fit. 13

14 Figure 2: Adjusted translation cost to English and per capita GDP.5 ISL log avg. pair rate net of source f.e. 0.5 ETH HTI TJK NPL BGD PSE AFG KHM ZAF CYP DZA FIN DNK NLD SWE GTM MLT JPN AUT PHL MKD KOR DEU ARE MNE SAU FRA YEM MYS IND ALB PRI ITA IDN EST ISR BIH MEX LBY LBN BOL ARM TUR KAZ SGP BRA EGY SYR MAR JOR LTU GEO CUBSRBGR POL URY SVN KWT GRC ESP COL SUR ROU CZE IRN MNG CHN AZE CRI VEN ARG LVA HUN CHL LAO VNM RUS SVK TWN PRT HNDLKAUKR ECU SDN IRQTUN DOM PERPAN UZB MDA SLV BLR NIC HRV QAT log country GDP per capita Vertical axis: log of the average translation rates between a country s most widely spoken language and English (undirected rates, i.e. combining rates where English is the source or the target language), net of source language fixed effects, plotted for countries for which English is not the most widely spoken language. English-Spanish rates for translators located in the United States are normalized to zero. Black line is linear fit. 14

15 Table 1: Trade and language barriers. Gravity linear regression on positive trade flows Adj. translation rate (fractional) (1) (2) (3) Log trade do > 0 Log trade do > 0 Log trade do > [0.301] Translation rate observed (fractional) [0.075] Adj. translation rate (top pair) [0.242] Translation rate observed (top pair) [0.060] Log fraction common language [0.005] [0.005] No common language [0.064] [0.064] Common official language [0.049] [0.055] [0.055] Log distance [0.021] [0.022] [0.022] Contiguity [0.102] [0.100] [0.100] Colonial tie (ever) [0.128] [0.130] [0.130] Colonial tie (after 1945) [0.169] [0.169] [0.169] Common colonizer (after 1945) [0.064] [0.065] [0.065] Exporter and importer f.e. Yes Yes Yes Observations R Standard errors in brackets p < 0.05, p < 0.01, p <

16 Table 2: Trade and language barriers. Gravity exponential regression on all trade flows Adj. translation rate (fractional) (1) (2) (3) Trade do Trade do Trade do [0.266] Translation rate observed (fractional) [0.094] Adj. translation rate (top pair) [0.235] Translation rate observed (top pair) [0.073] Log fraction common language [0.008] [0.008] No common language [0.128] [0.126] Common official language [0.109] [0.110] [0.107] Log distance [0.038] [0.039] [0.038] Contiguity [0.115] [0.108] [0.110] Colonial tie (ever) [0.135] [0.144] [0.143] Colonial tie (after 1945) [0.233] [0.262] [0.258] Common colonizer (after 1945) [0.163] [0.162] [0.162] Log gdp and remoteness (o & d) Yes Yes Yes Observations Log lik e e e+09 Standard errors in brackets p < 0.05, p < 0.01, p <

17 Table 3: Two-stage estimates of language barrier effect Second stage: (1) (2) Log trade do > 0 Trade do Adj. translation rate (top pair) [0.263] [0.283] Log fraction common language [0.012] [0.011] No common language [0.157] [0.155] Common official language [0.180] [0.192] Log distance [0.068] [0.044] Contiguity [0.196] [0.108] Colonial tie (ever) [0.276] [0.178] Colonial tie (after 1945) [0.288] [0.282] Common colonizer (after 1945) [0.410] [0.406] Log gdp and remoteness (o & d) Yes Yes First stage: Log ethnolinguistic overlap (0.021) (0.021) Observations R Standard errors in parentheses for coefficient on adjusted translation rate (top pair) is block bootstrapped at the country-pair level, and for first stage coefficient on log ethnolinguistic overlap is bootstrapped. Conventional standard errors in brackets are preliminary. p < 0.05, p < 0.01, p <

18 Appendix A Data sources Translation rates: Baseline results are from average rates by language pair and country, downloaded from translatorscafe.com on March 6, 2013 and discussed in the main text. Rates are included and labeled as observed if there are at least five translators present in a directed language-pair and country. Trade flows: Trade flows for 2011 at the HS6 level are from the BACI dataset provided by CEPII, which are in turn based on United Nations Comtrade data. Gravity covariates: Distance, common official language and colonial link data are from the CEPII Gravity Dataset. Ethnolinguistic data: Counts of population by language group within country are from Ethnologue, 16 th Edition. R & D and advertising intensity: From Kugler and Verhoogen (2012), the ratio of advertising plus research and development expenditures to total sales, from the U.S. Federal Trade Commission (FTC) 1975 Line of Business Survey. Measures for ISIC 4- digit rev. 2 classification concorded to HS6. Measure of horizontal differentiation: Classification due to Rauch (1999). SITC 4- digit industries concorded to HS6. Appendix B Estimation with a censored endogenous regressor I follow the method of Chernozhukov, Rigobon, and Stoker (2010) for estimation of a linear conditional mean model with a bound-censored and endogenous regressor. To describe the estimation approach, assume ln X = βl + D δ + U (2) L = Z π + V (3) U = γv + ε (4) where ε is mean independent of (V, L ) and V is median independent of Z. The dependent variable X stands for exports and L is the uncensored language skill premium, which is endogenous when γ = 0. D is a vector of standard gravity regressors such as distance, and Z is a vector of instruments that includes D. We do not observe L for all pairs of languages, so for all unobserved pairs I set the language skill premium at its highest observed value L and assume that this is an upper censoring threshold such that an observed, censored language skill premium L is given by 18

19 { L if L L = < L L otherwise (5) I estimate equation (3) in a first stage by censored quantile regression, employing the method of Chernozhukov and Hong (2002). Residuals from this first stage can be used as a control function for inclusion in a second stage, which can be estimated on the subsample above the censoring threshold. Construction of the control function for a linear conditional mean model follows directly. Applying the control function approach to the exponential conditional mean model for the gravity equation (as in Silva and Tenreyro, 2006) requires additional assumptions. I modify equation (2) to [ E X L, D, V ] ( = exp βl + D δ + γv ). (6) Inclusion of a control function γv in equation (2) is a stronger functional form assumption, for which a sufficient condition is joint normality of (U, V ). As censored quantile regressions are difficult to estimate with fixed effects, I replace exporter and importer fixed effects with importer and exporter gross domestic products and remoteness measures, following Baldwin and Harrigan (2011). Alternative gravity estimation methods aimed at removing country fixed effects (e.g. tetrads, see Hallak (2006) and Head and Mayer (2013)) are unsuitable in this context because they pass censoring points through non-linear functions. I compute standard errors by bootstrapping country pair observations across both stages. I do not resample translation rate data, as I view the fixed effect regressions to net source language effects and countryspecific wages from nominal translation rates as a data construction step. 19