AdobeStock_106108248.jpeg
VoxEU Column International trade Productivity and Innovation

The potential impact of machine translation on foreign trade – caution, please

Artificial intelligence has made spectacular progress in recent years. One particular source of high expectations is automatic translation and whether it will finally bring about the long-predicted death of distance in trade. This column examines the impact of a common language on bilateral trade and finds that the net result of reducing linguistic frictions with a set of trading partners is not apparent.The potential impact of machine translation on foreign trade remains up in the air.

Artificial intelligence has made spectacular progress in recent years. Machines now can modify their path to objectives based on the information they collect, rather than on set algorithms. They can learn. This may have profound effects on the future of mankind. Will it finally bring about the long-predicted death of distance in trade?

One particular source of high expectations is the advance of automatic translation. “Machine learning is tearing down language barriers. What does this mean for trade?”, asks Richard Baldwin in a recent Vox column (Baldwin 2018). In response, we reply: ‘Don’t hold your breath’.

Baldwin points to the sole study thus far that is right on the topic. In a careful statistical work, Brynjolfsson et al. (2018) show that US firms have increased their sales to Latin America via a particular digital platform, eBay, by 10.9% since the introduction of improved machine translation in 2014. On close examination, though, the study does not say whether the increase in sales to Latin America by certain US firms via eBay was at the expense of sales by other US firms to the same destination. In so far as total sales to Latin America did rise, it may have been at the expense of US exports to other parts of the world.

Are these mere quibbles? We think not. Brynjolfsson et al. and Baldwin both point to the gravity model to support their claim of potentially large effects on world trade. Yet, the usual tests of this model strictly concern the choice of different foreign partners in trade. The issue of openness is entirely absorbed by country-specific fixed effects.

In effect, the whole question of the potential impact of machine translation on foreign trade remains up in the air. While multilateral trade is the central issue, since it concerns the invasion of foreign trade into domestic, examining the potential impact of a common language on bilateral trade is a good way to start.

Common language effects on bilateral trade

Machine translation will bear on bilateral trade strictly via the ability to communicate, whereas linguistic effects on trade largely come through ethnicity and people’s wish to associate and bond with similar others, and perhaps also through similarity of tastes. Contemporary discussion on the impact of language on trade, dating back only 25 years or so, began with stress on the ethnolinguistic aspects. It is only later that communicative benefits came to share the spotlight.

As far as these latter benefits are concerned, the distinction between direct and indirect communication is important. Machine translation will undoubtedly improve indirect communication. But what about direct communication? Baldwin is optimistic about future inroads of machine translation in this area as well. Machine translation, he points out, already promotes matchmaking between people without a common language and the ability to converse to some extent. But how far will such progress go? Will machine translation ever improve enough to induce trading firms to lose interest in hiring native speakers and financing the learning of foreign languages by existing staff (Hagen et al. 2006)? And what about the role of homophily and heterophily in direct exchange?

In fact, the empirical literature has a great deal to say on these questions. When a common official language enters with a common spoken language in empirical tests of the gravity model, the official language essentially reflects the role of indirect communication. Its coefficient signals the extent to which public support of a language facilitates the ability to obtain messages in that language by those who do not know it, through translation. Thus, it informs us of the extent to which artificial translation can diminish language barriers to trade.

The coefficient of common official language, in this case, is about a third of the total (Melitz and Toubal 2014). Specifically, in worldwide samples, common spoken and official language together have a semi-elasticity of influence on bilateral trade of about one, of which common official language accounts for around 30-35%. (The oft-cited figure of 0.5 for the total effect of a common language occurs when common official language serves alone.)

If those numbers seem suspiciously high, the reason is likely to be a failure to keep in mind that they strictly concern the choices of alternative trade partners. If a new bakery appears closer to your residence, you may switch all your trade to this new one without necessarily consuming more bread. So far as this analogy applies, the figures are entirely reasonable.

How likely is it that automatic translation will cut down the impact of a foreign language by the full third? For the moment, we know little of the answer.

What about the other two-thirds of a common language’s influence on trade, the part coming through common spoken language? Here, it pays to distinguish between homogeneous, listed, and differentiated goods in the Rauch classification. Homogeneous goods are those quoted on commodity exchanges. In their case, we could well imagine that the market would be so impersonal and the required information so small that a common language would not matter at all. Indeed, common official language plays no role. There is thus no sign that translation has a separate importance.

Correspondingly, a common religion has no role for homogeneous goods either, though it does for listed and differentiated goods. Therefore, personal affinities seem not to be at play. Yet, perhaps incomprehensibly at first, spoken language is, if anything, more important for homogeneous goods than for the other two types of goods.

The clue lies in the importance of a common legal system. Evidently, there remain concerns regarding litigation. Contracts may be broken, shipments may arrive late or with spoiled or missing goods, it may be necessary to go to court abroad. If so, it’s better to know the foreign language of the seller. Will improved machine translation diminish this interest in direct communication?

Let us consider next the contrasting case of differentiated goods, where the buyer knows the particular supplier. Official language accounts for over a third of the total influence of common language in this case. There are three indications that automatic translation will not go much further in diminishing the role of a common language in the rest. First, of the three Rauch classifications, this is the one where the separate impact of a common native language is important enough to be distinguishable from that of a common spoken language. Evidently, personal affinities matter, quite apart from the ability to converse. Second, common religion has a greater influence in this classification than in either of the other two, another clear sign that a sense of fellowship is a factor.

Third, we present in Melitz and Toubal (2019) new evidence that somatic distance between national populations, as measured by average height, colour hair, and head shape (ratios of width to length) have a marked influence on bilateral trade, apart from linguistic, religious, and legal influences. This evidence is limited to Europe. But human physical appearances are distributed far more narrowly there than the world as a whole, and we therefore argue that the evidence might be applied worldwide. In these tests, somatic distance notably cuts into the effect of common native language, the single measure of common language present. Surely, therefore, in the absence of somatic distance, common spoken language partly reflects its influence. This aspect of a common language probably relates to the importance of homophily in social behaviour that sociologists find everywhere. 

Common language effects on aggregate trade

Next, let us turn to the central issue of linguistic influences on aggregate trade, or the ratio of domestic to aggregate foreign trade. Two factors would show how uncertain the issue is. In their celebrated paper, Anderson and van Wincoop (2003) insisted on viewing the influences of all bilateral variables (as opposed to country-specific ones) in gravity models as the ratios of their effect on bilateral trade relative to multilateral trade. The subsequent literature effectively failed to adhere to this point by merely introducing fixed country effects, which meant recognising multilateral trade resistance but setting it aside.

One result inspired by Anderson and van Wincoop indicates the influence of common language on multilateral trade. In Melitz (2008), I estimate both the influence of common language on trade with fixed country effects and with a linear approximation to Anderson and van Wincoop’s tailor-made method of estimating ratios of bilateral to multilateral trade (or ‘trade resistance’) proposed by Baier and Bergstrand (2009). My common language measure combines official and spoken language.

For bilateral trade with country-fixed effects, the semi-elasticity of a common language’s influence is 0.95; for the ratio of bilateral to multilateral trade, it is 0.81. This lower estimate says that the introduction of multilateral trade resistance lowers the impact of higher common language on bilateral trade. Greater common language lowers foreign trade in the aggregate. This is admittedly a counterintuitive result that requires further confirmation, but we wish to point out why it is possible. 

Consider the impact of a reduction in linguistic frictions with a set of trading partners. The home country will switch its trade to this set from the rest. In addition to this switch, home firms may also switch a part of their domestic sales to them. On the opposite side, though, sales to the rest of the trading partners suffer. We can then imagine that because of the fixed costs of trade abroad, home firms would find the home market more attractive relative to exports. This turn toward home sales need not sway the aggregate, but it could. The essential point is that there is a countervailing effect; the net result is not transparent.

A final linguistic issue to keep in mind is linguistic diversity at home. Such diversity reduces the advantage of favouring the home market to avoid the costs of a foreign language because those costs come in home trade as well. In Melitz (2008), I find that linguistic diversity at home indeed increases openness.

Since detailed evidence is lacking, let us rely on a plausible example. Switzerland is a small, open country with many German- and French-speaking residents who do not speak the other language. Suppose that automatic translation improves enough to eliminate all problems of communication between German and French speakers. German speakers will then purchase more from French firms at the expense of German firms on the other side of the border.  French speakers will buy more from German firms at the expense of French firms on the other side of the border. It is easy to imagine that domestic trade would rise relative to foreign trade.

Is it true then that, as Richard Baldwin writes, “as machine learning tears down the language barrier between major languages, world trade flows should rise – a lot”?

References

Anderson, J, and E van Wincoop (2003), “Gravity with gravitas: A solution to the border problem”, American Economic Review 93: 170-92.

Baier, S, and J Bergstrand (2009), “Bonus vetus OLS: A simple method for approximating trade-cost effects using the gravity-equation”, Journal of International Economics 77(1): 77-85.

Baldwin, R (2018), “Machine learning is tearing down language barriers. What does this mean for trade”, VoxEU.org, 21 September.

Brynjolfsson, E, X Hui and M Liu (2018), “Does machine translation affect international trade? Evidence from a large digital platform”, NBER Working Paper 24917; forthcoming in Management Science. 

Hagen, S, J Foreman-Peck, S Davila-Philippon, B Nordgren and S Hagen (2006), “ELAN: Effects in the European economy of shortages of foreign language skills in enterprise”, CILT, The National Centre for Languages.

Melitz, J (2008), “Language and foreign trade”, European Economic Review 52(4): 667-99.

Melitz, J, and F Toubal (2014), “Native language, spoken language, translation and trade”, Journal of International Economics 93(2): 351-63.

Melitz, J, and F Toubal (2019), “Somatic distance, trust and trade”, Review of International Economics 27(3): 786-802.

2,310 Reads