For two years, I worked with the talented team at Nuance designing chatbots for Fortune 500 companies like Esurance, AT&T, Verizon, and more. 

While many of my skills from previous roles have been transferrable, I've been developing my skillset in working with statistical language models to make conversational interfaces that employ language successfully to reduce comprehension errors, and therefore user frustration. 

Problem: 

Often, clients will come to Nuance with a rough idea of what content they'd like their Virtual Assistants (VAs) to cover. We'll pull in data from their IVRs, if they have them with us, other live-chat applications, and other sources to determine their most high-volume content. From there, we need to make sense of this data and organize it in a way that not only will be clear and usable for users, but fits into a statistical language model to improve comprehension. 

Process: 

I work closely with Speech Scientists, Developers, and Clients to determine what content needs to be covered by Virtual Assistants. Once content is confirmed, it goes through a very detailed process of information architecting to ensure that we are designing a grammar that is optimized for users.

We divide content areas out into larger content topics, and divide by subptobics and concepts within those larger buckets. Each of these help inform our complex decision tree flows, and ensure that we have intents that aren't overlapping and cause grammar confusion. 

An example of this type of content breakdown can be found here. 

Feedback: 

I train clients on all of this complicated intertwining of content and interaction design, teaching them to think of the content as a complex web of a conversation rather than a 1-1 call and response system. 

Results: 

I evaluate the success of these Virtual Assistants over time to ensure the grammar is still working successfully. In cases where users are not asking certain questions in high volume, there are large numbers of incomprehension, or otherwise - those are ripe moments for redesign and a review of how the structure of the grammar was made. 

Conclusions:

I'm growing a great deal on this team, and better understanding how to construct complex conversations with functionality that brings in variable content, has complex decision trees, and working with larger profile clients. I will continue to update this space as I take on new client challenges.