One of our Software Engineers, Nemanja, and his colleagues in Belgrade recently created a chatbot. Their bot is called Zubot, short for Zuhlke Bot, and it has quickly become one our most beloved employees since it is so much to talk to. Zubot is mostly active at career fairs where it can share information about Zuhlke as an employer to interested talent, while also contributing some humour and chit chat.
What type of person is Zubot?
It has a very good sense of humour, is entertaining and very well mannered. It will always say “Thank you”. On the other hand, it recognizes if someone is rude. Then it will answer adequately.
Who created Zubot?
In the beginning of this year, Zuhlke Serbia established internal focus groups during which employees could focus on innovating outside of our clients projects. During one of those, three colleagues and I started working with chatbots.
How did the idea of creating a bot for career fairs evolve?
Initially we wanted to create a food ordering bot, but it turns out we already an app for that at Zuhlke. Then our HR colleagues came up with the idea of creating a bot for career fairs. The idea was that it would be able to answer simple questions and provide a playful way for talent to interact with Zuhlke as a company. With an upcoming career fair in mind, we had five months to create the bot from scratch, to test it and get it running.
How is it done technically?
When we recieve input from usage, we forward it to our natural language processing service. It deduces the meaning of the sentence and provides a proper answer. We use a combination of two services. One is Wit.ai and the other is Google NLP API. Input is very valuable to us because we use it to train the model and it improves over time. On every question we receive, we put a label, which helps to figure out the meaning of the sentence and return what is known as entity. This is how our logic performs. Although both services have performed very well, we are currently testing another library called NLTK. It is one of the most popular natural language processing libraries.
How does Zubot learn and improve itself?
We tested it internally with our colleagues and improved it through their feedback. These tests took us a while, but the user feedback is really valuable. There you see the actual questions asked. This Iteration was very important and made the bot as good at it is now.
How did people react to the bot at the job fair?
The bot is on our iPads and the students were excited and loved it. I was there and could observe how people interacted with our bot like it were a real person. Conversation is a natural thing for humans, even with a bot. There were many funny reactions and it was rewarding for me to see them.
Will there be more Zuhlke bots in the future?
In the upcoming months, we will be focusing on improving the natural language understanding. We are not fully satisfied with our model. We see gaps, which we want to improve. This field is very broad and contains many techniques. We will do a lot of testing until we find the right one. One big drawback of the current version is that the bot is not aware of context. We want to change this in a future version.
„Conversation is a natural thing for humans, even with a bot.”
What is the best thing about building a bot?
People’s reactions. It is still software but with a real flavour of natural language, people interact in a completely different way than with other apps, which is very interesting to observe.