Prioritizing urgent customer requests, regardless of language: Yandex Taxi’s story

Yandex Taxi is one of the world’s largest ride hailing services, operating in 19 countries. What also sets Yandex Taxi apart is its level of technological advancement: machine learning is used throughout the entire order cycle, including the support service.

The support service receives a steady stream of requests every day, which a ML model then prioritizes to ensure that the most urgent ones receive an immediate response. The model is only trained for the Russian language, however. By bringing in Yandex Translate, the model is able to understand other languages, making it possible for the support service to quickly respond to all requests, no matter what language they are in.

The Yandex Taxi service connects drivers and taxi companies around the world. Every week, the service support service receives hundreds of thousands of requests in different languages. Bots are used to quickly process these significant volumes of information. They answer simple standard questions and call in support staff if the problem requires a response from a human employee.

But what about emergency situations, when a passenger needs an immediate response from a support representative?

The Yandex Taxi team faced the technological challenge of teaching the algorithm to respond to these kinds of situations, without the user’s language getting in the way.

The solution was to use machine translation within the ML model. The translation is carried out by Yandex Translate, which supports more than 95 languages. This application is called Machine Translation for Machine Learning. Machine translation is not used for people, but for the machines themselves, which can process multilingual content based on this translation.

First, Yandex Taxi compiled a tree of topics of customer queries in Russian and taught the ML model to navigate them. There were about two hundred possible problems: with the trip (the driver never came), with the mobile application (I can’t link my card), with the vehicle (the car is dirty), with forgotten items, etc.

Here’s an example customer query: “Paliko juodą žiebtuvėlis Zippo. Tai dovana. Būčiau dėkingas, jei aš atgal” (Lithuanian) — “I forgot a black Zippo lighter. It was a gift. I would be grateful to get it back”.

Special attention was paid to messages that require an immediate response: the ML model learned to recognize and prioritize them for the support team.

Via Yandex Translate, the ML model could understand customer queries in languages other than Russian. All messages in other languages are automatically translated into Russian, processed by the ML model, and assigned the appropriate priority. They further trained the model based on accumulated support data, with an emphasis on teaching it to recognize and identify transliteration.

Training the ML model to work in parallel in other languages would have been very labor-intensive and inefficient. This would have involved In this case, we would have reproduce the process for all languages, i.e. search and gathering of relevant data in other languages, experiments and selection of the best models, quality assessment (especially difficult when speakers of various languages are required), deployment, quality monitoring, updating, etc.

Machine translation turned out to be the ideal solution, thanks to the speed and complete autonomy it offered. One machine’s interaction with another brings ease into people’s lives, and what was once expected to take place in the future is now here in the present.

Now that the support service can instantly understand customer queries in 95+ languages, Yandex Taxi has improved their service quality, and can respond to user queries much faster.

If something goes wrong, customers can always contact the support service, no matter what question, problem, or unpleasant situation might be. The most urgent cases are processed first. And thanks to the integration with Yandex Translate, the original language of the request makes no difference.