Research and development project with Laurea University of Applied Sciences and Hyvinvoiva Terveydenhuolto
We were recently privileged to take part in a research and development project with Laurea University of Applied Sciences and Hyvinvoiva Terveydenhuolto (https://hyvinvoivaterveydenhuolto.fi/), roughly translated “Wellbeing to Healthcare”.
The aim of this EU-funded project is to respond to the urgent need of nurses all over Finland to find ways to deal with their challenging workload.
“We wanted to explore the asset-based and solution-focused approaches with our chat bot as an alternative to the more common problem-focused approach. Instead of identifying a person’s problems and the possible causes resulting in the difficulties they are facing, in our approach the main attention is directed to their desires, successes, resources and the reasons for the progress that has already taken place. We also expected this sort of an approach to be easier for a chat bot, because we could train it to ask the right questions that help an individual figure things out for themselves, instead of training complex pathways that require a lot of guesswork and probably end up in a “sorry, I don’t understand” situation anyway.”
– Hilkka Lydén, project manager
The project was a welcome challenge, because it pushed us at Front AI to think about completely novel ways of solving the problem using our technology.
Happily, Laurea had carried out a lot of user testing beforehand. We could proceed with lots of real user conversation data, which undoubtedly contributed to the good end results.
Analysing Laurea’s findings and test conversations led us to a formalisation of the conversation process. The process could be broken down into five separate phases, with three levels of questions in each phase. Implementing this in our conversational AI platform was made simple by its programmatic nature.
Instead of attempting to understand the user to a detailed level, which is the usual aim of a conversational AI project, the goal was to provide an experience where the entire point is the user thinking about and answering questions presented to them.
This meant that the logic of how we normally build projects was turned upside down. As opposed to trying to analyse and understand the user’s message, we built a process that would always return the user to the next appropriate question or phase in the conversation. The bot is the one asking the questions, not the user.
An interesting aspect that follows is that Unknown predictions, where the AI fails to predict to an intent, are the most common type of prediction and completely expected. The Unknown prediction is simply understood to be the user’s answer to the question, and we move on to the next question.
On the other hand, actual predictions and model training were restricted to cases where we knew the bot had to react in some specific way, such as pointing to external resources at the mention of suicidal thoughts or self-harm, or when the user says “I don’t know”, in which case we encourage them to try again and do not count the question as answered.
Within a very short period of time, we were able to build a conversational process that feels genuinely helpful and valuable. Further developing the model to better understand the user could enable us to make the process more responsive by setting and examining variables that determine which questions are asked.
Another aspect that was considered, though not implemented in the first sprint of this project, is learning from older projects. ELIZA (https://en.wikipedia.org/wiki/ELIZA) was a groundbreaking “therapy bot” in the 1960s, whose principles of mirroring back the user’s utterances could be used to elevate the experience to a whole new level.