Chatting it up with the Tech Team
By Jessica Wray - January 19, 2018
I joined Sciensio four months ago as an AI Interaction Designer and have learned an immense amount about what it takes to build a successful chatbot from the conversation side. My team has built chatbots for an interesting range of events including trade shows, charity auctions, car test drive events, and corporate meetings, as well as some non-events. I’ve learned a lot about what makes a bot conversation successful and what doesn’t. However, being a former software developer, I was curious about the differences between developing for chatbots as opposed to applications. So, I asked our development team what they had learned in the last year and this is what they had to say.
The Hard Part
I was ready for them to get right down to the ins and outs of our backend systems. I was surprised when they told me the number one thing they had learned was the technology wasn’t the hard part. It’s the conversation! To that end, a lot of the code our team writes focuses on improving the user’s experience.
One way they’ve done that is providing a custom initial greeting based on what we know about the user. Our code takes into account what channel they are using to interact with our chatbot as well as how the user started the conversation. This part of our backend system enables us to deliver a personalized experience for the user.
Another way our developers have helped to deliver a better conversation is by customizing the pace so the user has time to absorb it. We do this by breaking up the conversation into readable bubbles and timing the way each message is delivered. These are a few of the ways our developers have made it easier for us on the AI Interaction Design team to be more efficient, build better conversations and leverage the lessons we learn from one chatbot to the next.
Every Chatbot is Learning Lab
This was the most exciting part of the discussion we had. Every chatbot meant we had another opportunity to watch how users interacted with it and we were able to immediately learn and iterate.
For example, we thought if we told the user to get started by greeting the bot with “Hi” they would. Nope! Many users skipped the formality and went straight to asking the bot a question. Since the bot was expecting a simple “Hi” it returned it’s standard greeting which didn’t address the user’s question. This doesn’t make for a good start to the conversation and the user was left confused. This is where our dev team stepped in and built a smarter initial greeting handler that improves the user’s boarding experience.
This is one of the many lessons we’ve learned. You can only imagine how much invaluable data we collected each and every time. After every bot implementation the things we observed and learned were incorporated into our core platform. Our dev team was able to make many of these changes quickly so the next bot we delivered was more successful each time. As a result of these fast changes, the differences between our bots at the beginning of the year to those at the end of the year are night and day.
Integrations are Key
Conversation design is definitely important and regardless of what platform you use to write your bot conversations, it’s still only one piece of delivering a robust chatbot. There are an overwhelming amount of systems out there. Figuring out which ones to integrate your chatbot with is as important as designing the conversation itself and this is where our developers focused most of their time and efforts. First, they needed to figure out what integrations made sense and then they needed to make sure those integrations helped our bots be smarter. By choosing the right integrations we were able to deliver a richer, more personalized experience for our users.
We integrated with systems that would give us data specific to the user which enabled us to deliver a more personalized conversation. We were also able to enrich the user experience by integrating with systems such as Google Maps to provide driving directions. Perhaps the most important way we improved the conversation was by recognizing we couldn’t rely solely on Natural Language Processing (NLP). While NLP has definitely improved since we started, it still needs support to deliver a robust chatbot. It struggles with spelling mistakes, abbreviations and the thousands of ways a question can be asked and that’s just the beginning. Our dev team focused on integrating with services such as spell check, language translation and sentiment analysis to extend our NLP and greatly improve the conversation.
The exciting part for me is that it’s a new year full of new chatbots. We recently all took a day together to make some big plans for our chatbot platform, both technically and non-technically, and we are looking forward to launching and learning even more in 2018! Stay tuned…