Not all of our intelligence is artificial: meet the Caravelo bot trainers

In 2018 chatbots have become commonplace. Their ubiquity underlines just how easy it is for automated agents to enable greater customer service and retailing. The AI behind bots is open to everyone and the difference that exists between good bots and not so good bots is not the machine learning, but the thought models and training that are behind it.

Here at Caravelo we have a great team of bot trainers whose job it is to take the ambition of our customers and bring it to life in conversation form. What better way for our bot trainers to share their insights than through a conversation… and if you’d like to have a conversation about bots, be in touch!

So, why do bots need trainers in the background?

Andrew: Well, Caravelo bots are built on the principal of supervised learning which means that the machine learning algorithms need human training to get smarter. This ensures that bots get brighter and only learn about what’s relevant to their job.

Onur: If a bot is trained sloppy, all the great work done for development, design and training of a bot becomes useless.

What does it take to be a great bot trainer?

Iina: I think you need to be both imaginative and systematic. You have the creative part of the job when designing the bot persona including their tone of voice, then on the other hand the planning and creation of the flows and execution of the training. It’s good to know some coding, and also problem-solving skills come in handy, too.

Andrew: Perseverance and determination. It takes time to train a bot to a good level, where it’s actually useful to a user/customer. You have to trust the technology, rely on the algorithms getting things right, and test and re-test and fine-tune. The results are not immediate, but worth it when you see the reactions of real end-users.

What does a typical day in the life of a bot trainer look like?

Iina : A vital part of our work is to check the interactions that users have had with our chatbots. That way we can see when the bots have failed to understand the user or misunderstood what the user was asking. This real-life insight gives us a chance to train our bots to be better every day, to keep them constantly improving.

Onur: Truly, training is only one small aspect of our daily work! We can talk about two main tasks we fulfill at Caravelo: (i) we contribute to the design and creation process of a new bot, and (ii) we are responsible for the maintenance and improving of any live bot connected into the platform.

In the first phase we define the scope, priorities and personality of the bot according to the needs and expectations of the client. While our dev team focuses on the integration of the backend functions, we focus on the conversational and linguistic elements. We list the assigned “intents”, map the relevant content accordingly and build a conversational structure. Then starts the training process: We have to match the intents with variations of related questions and feed the bot with a carefully selected set of utterances, which are supposed to be recognized and trigger correct answers. As we grow, adding those intents gets easier – we now have over 700.

Next phase is a long process of testing for problem solving purposes and fine-tuning the dialogues. Ultimately, our days are challenging, but it never gets boring. And it’s always amazing to watch a bot going live which you have created and trained!

What does your engagement look like with the airline?

Iina: This varies from airline to airline. Some of them like to be very hands-on in the process, helping to mold the personality of the chatbot and making suggestions on the content of the bot. Others then give us a lot of freedom to create on their behalf.

Andrew: Our airline partners embrace the use of this technology so it makes our relationship easier, as we both want our chatbots to succeed. The fact that this is now what the future will look like makes it exciting too. It’s a constant process of improvements which the airlines can see on a daily basis.

Tell me something surprising about bot training?

Iina: The way people communicate with the bots always keeps on amazing me. They can ask the bot out on a date, joke with them, swear at them… and sometimes tell touching things such as that they are feeling lonely. How do you train your bot to handle this vast scale of human emotions? It’s really a surprising responsibility you get as a bot trainer.

Elena: People sometimes write long letters to the bots as if they would talk to their friends or, on quite the opposite, as if they feel much more comfortable talking to a robot than to a human.

Onur: I am always surprised about the magic of randomness and “positive” failures! Somebody asks “What is the meaning of love?” and your bot responds randomly or “mistakenly” with the perfect answer.

Apart from AI and human training, what are the other ingredients in building great chatbots?

Onur: Chatbot design is much more than the mere sum of bot development, UX/UI design and bot training. I would say chatbot design is closer to script-writing for a film, than to web design, for example. It involves a deep understanding and mediation on human personality, human language and human situation, as we design and create a virtual person.

Andrew: Research is important. The conversational experience is new and growing, and there are many experiences out right now. It’s always great to test other bots, seeing the different takes and styles out there. Also combining other design and writing ideas from different user interfaces on all kinds of platforms, as well as creative writing and much more. This gives you a better perspective on what’s out there, and how it applies to your bots.

Iina: What you really need to take into consideration is the user experience. They come to talk to bot in different situations, maybe planning a dream holiday or possibly extremely stressed as they have missed their flight and don’t know how they will get to fly home. How do you make their interaction with the chatbot as effortless as possible for these scenarios?

How do you see the future of chatbot training?

Iina: Exciting! Already now we can see how people are used to communicating in messaging apps by sending screenshots, photos, stickers and of course emojis. The emojis we can train the bot to recognize, but with the other things it gets more tricky. And then there’s of course the voice recognition.

Elena: I believe chatbot will definitely evolve from “text” to “voice”. We get busier and busier everyday, and talking is always much faster than writing…it would be a very challenging training as everyone writes in an unique manner but when it comes to talk… that’s another story.

Onur : What is also exciting is how fast artificial intelligence is progressing. Semantic analysis, context analysis and emotional analysis are becoming more relevant. We already have bots who are able to “read your mind”, your emotional mood and respond accordingly. Very advanced bots will be able to learn from their own experiences and even from other bots, and they will be able to answer questions which they are not trained for!