How Do You Train a Chatbot?

We see chatbots popping up more and more on all sorts of websites. Chatbots that use machine learning need to be fed a lot of training data so that they can learn to understand and answer users’ questions. This data can come from pre-existing customer service emails, call logs and FAQs, but that’s often not enough. That’s where a language service provider comes in. We can write hundreds or even thousands of utterances that reflect what your target user might ask the chatbot and how they might ask it.

We Recruited 120 People and Asked a Lot of Questions

This spring, OXO was hired to provide Quebecois French training data for a white-label customer service chatbot being developed by a tech giant. The project involved a total of 4,500 customer service scenarios for different industries. A typical scenario is something like, “You just bought a plane ticket and you want to know if pets are allowed on the airline” or “You want to change cellphone providers but you’re worried about losing your number.” For each scenario, two people had to each come up with 55 ways of asking the same thing using different tags and keywords and not repeating sentences. And they had about an hour do so. The client will then use these questions as chatbot training data.

This was a new challenge for OXO. To complete the project within a one-month deadline, we had to recruit and test 120 people. People who were available to start right away, who were able to write in French without mistakes, and who had the creativity to come up 55 ways of asking the same question on the spot, which not everyone is able to do.

Beyond this creative challenge, the main hurdle we encountered for this project was of course the human resources angle—we had to recruit, test and manage a lot of people. It’s something we’re used to doing, but never on this scale. Our Human Relations Manager, Kevin Lavoie, reached out to our large network of collaborators as well as local universities and encouraged staff to spread the message among their personal contacts. The result: we were able to recruit over a hundred people within two weeks and provide a good source of income for teachers, tutors, translators and recent grads who were unemployed due to the COVID-19 pandemic. We completed the project within 1 month, and the client was super pleased with the quality.

Expanding the Definition of Language Services for the 21^st Century

We have to admit that, before we were approached for this project, it’s not a service we had thought of offering (despite “Innovation” being in our name!). Yet it makes total sense—we already have the robust vendor management structure and linguist expertise in place. In fact, many of the world’s largest translation agencies are already pushing machine learning services. It’s time for the littler guys like us to catch up. The artificial intelligence industry is only going to keep growing, especially here in Montreal. The demand is certainly there, and there’s room on the playing field for OXO’s customer-oriented and quality-conscious approach.

There’s concern in certain circles of the localization industry about AI making our jobs obsolete. The demand for chatbot training data is just another example of the ways in which technology—far from making us redundant—is creating new opportunities for people with project management or language skills.

OXO’s Machine Learning Services

Now that OXO has successfully completed this first chatbot training project, we’re excited to expand our services in this direction, and we’re fully equipped to take on more machine learning mandates.

Here’s what we can offer:

• Highly experienced project managers and vendor managers
• A large network of resources with strong language skills and an interest in creating chatbot datasets
• Localization for many languages and regions, including:
  • North American English
  • Canadian French
  • Brazilian Portuguese
  • Latin American Spanish
• Headquarters in Montreal, a global hub for AI

If you’re interested in learning more about this service, contact us today.