Founded in March 2006 and launched in July 2006, Twitter is a real-time information service on which people post ideas, comments and news in 140 characters or less. Twitter brings users closer to the topics, events and people they care most about around the world. Twitter is available globally in 33 languages, with 140 million active users and 400 million tweets per day. Based in San Francisco, Gaku Ueda is Twitter’s engineering manager in charge of making Twitter an even more global product by translating it into more of the world’s languages.
Thicke: You are the manager of the international engineering team. How did you come to this role?
Ueda: I have been working on globalization for more than 12 years. In my last job, I was the technical lead/manager for the Japanese market adoption of Google Maps and before that I was at Yahoo! I came to Twitter in November 2011 to start leading the engineering efforts around internationalization.
Thicke: How does language fit into Twitter’s vision?
Ueda: One of our core values is to reach every person on the planet. We care about everyone on earth; this is why we are also supporting minor languages. We recently covered Catalan and Basque. Again, we care about these languages.
Thicke: How many languages is Twitter offered in today?
Ueda: Twitter is offered in 33 languages today, and another 13 or so are coming. More than half of our traffic has been coming from outside the United States. And it’s growing. International is an important part of our growth.
Thicke: How do you manage so many languages? Are they all handled by volunteers?
Ueda: Yes, we are doubling down on community-based translation. Basically, we are relying on working with a community of volunteers. It’s been very successful.
Thicke: How do you define success?
Ueda: We define success in three ways. First of all, there is the translation quality. Because our volunteers are avid Twitter users, they understand the product before translating the string. They tend to come up with good translations. The turnaround time is also good: when we introduce a new feature, the volunteers respond quickly. Then there is the language coverage. It would be difficult to support a large number of languages if we were doing traditional translation. With volunteers we can cover more exotic languages.
Thicke: How does it affect your users when Twitter becomes available in their language?
Ueda: We started supporting Arabic-speaking countries in March of this year. Since then the user growth of Arabic users has been accelerating rapidly. Each month more people are using it, and I haven’t seen any sign of it slowing down. Before March the growth curve was very different. Language support is dramatically changing the landscape for these countries. Before localizing in right-to-left languages, Twitter was not that big in Saudi Arabia. People are using Twitter more as a result of localization.
Thicke: How do you decide on the “official” Twitter languages? Do you make a distinction between top tier languages, and second tier?
Ueda: We came up with a plan a while back to support 28 core languages. That was based on how much of the internet population we could cover. Now we are going beyond that — we are already supporting 33 and going toward 50 or so soon. We are switching to a self-service model to scale this up. If enough people want their own language version of Twitter, they can make it happen. Recent languages such as Basque and Catalan had enough volunteers who wanted it, so they made it happen.
Thicke: You have over 400,000 translation volunteers. That’s an amazing number. Do they all contribute the same amount?
Ueda: There is a low bar in terms of how much time is required to contribute. For example, when I see that some translations aren’t done, I can spend a couple of minutes and contribute to the Japanese version of Twitter. When we have a major launch, in the communities that are really active, thousands will respond. We also have a large number of translators helping with a steady flow of translations. They can contribute more and engage deeper by becoming moderators or participating in forums.
Thicke: How do you manage so many volunteers?
Ueda: One way we manage is our translation interface. http://translate.twttr.com has a dashboard of translation completion so you can see where translations are needed. Another way is that we have community managers who communicate with moderators about upcoming work. These are our two main communication channels. Also, we tweet.
Thicke: How is the work delivered back to you once it’s translated?
Ueda: It’s done via our website, and it immediately becomes available for the engineering team.
Thicke: How do you manage quality? I see a voting mechanism on your site. Is every sentence put to a vote?
Ueda: There are two ways of managing quality. One is moderation. We identified a subset of volunteers as moderators, so they have the right to approve translations. They are the gatekeepers of quality. Also, volunteer translators can vote for translations. They can vote for an existing translation or propose a new one.
Originally, when we were supporting 28 languages, moderators approved all the translations. Now, for the self-serve languages, the communities tend to be smaller and sometimes the string might not be approved. In that case, we have automatic moderation capabilities. If a translated string meets certain criteria, it may be automatically approved. It is machine based, so an algorithm decides.
We also have a reputation system for translators. The more good work you do, the more the system trusts you. If a string is not approved by the community within a certain amount of time, the system may approve the string based on a translator’s past record. If you are a brand new translator, we don’t know if you are good at translation, so we wait until we have enough confidence to approve the string. Voting is one of the inputs for the decisions. Approval goes to the moderator at the end. If the moderator doesn’t decide, then the algorithm decides.
Thicke: How long does the approval process take?
Ueda: It depends on how much we can trust a particular translator. If a string is translated by a trustworthy translator, it is approved very quickly.
Thicke: How much volume does an average translator handle?
Ueda: It varies by volunteer. Some translate a few strings, others come back every day — they like spending time to translate, review and vote. Some are putting serious time into it.
Thicke: How does the translation center distribute work? Is it push or pull, for example?
Ueda: In general it’s pull, but sometimes we push info to volunteers.
Thicke: What is the role of moderators, and how do you select them?
Ueda: We evaluate moderators by looking at past contributions. If volunteers have been contributing for a long time and are doing a good job of communicating with managers, we tend to choose them. We select the most active volunteers as moderators. We ask if they are interested, and if they are, we provide some training.
Thicke: I see you have a leaderboard with your top eight translators. But you have 400,000! How do you keep the rest of the volunteers motivated?
Ueda: There are a couple of additional things we do. One is that we have translator badges. The icon shows up in your profile page if you are a top 20% contributor for the translation community. Secondly, we have a learning center where you can play games about providing good translations. It’s like an online translation learning center. Also, on your translation center profile page you can see where you are — there is a dashboard for your contribution statistics.
Thicke: I see a vote for not only translations but also for the translator. How does this work?
Ueda: With the reputation system, you can evaluate the translation string or vote for particular translators. If someone is providing good translations, you can vote for that person too.
Thicke: How does mobile technology impact your plans?
Ueda: This is one of the areas for international expansion; the effort is mainly for users outside of the United States. Twitter is seriously supporting feature phones. Smartphones can do all the fancy things that a browser can do, while a feature phone’s browser feature is limited and the speed is slower.
Thicke: What languages are in the future for Twitter?
Ueda: We are talking about supporting 100 languages. That’s the next long-term goal for our team. We don’t have particular languages, but we are scaling the coverage of languages. We also want to improve the quality of translations. We want to make the product more local and natural for users outside the United States. So we are making sure that the Twitter product is tailored for individual countries for what is natural and useful for people in those countries.