[tdaltonc]

I'm curious about the middle-way chatbot you're building. Most chat systems are either "smart" scripts with intent recognition (RASA, Dialog Flow, etc) or anything-is-possible chaos (seq2seq, etc).

Can you explain more about how you're able to keep the conversation on track while leveraging the adaptability of a generative system?

[phowson83]

We're generally more in the smart scripts camp at present. The conversations are graph based, with each node being associated with code that can look at the state of the user and any input. It may be that the input is a set of fixed options, or that the input is freetext. Usually in the event of freetext we're classifying that freetext, looking at the state and using these two things to determine the output and next node to go to.

It's still early days with the product and to date we've generally avoided anything that is generative, as you say, it's hard to keep on track and is hard to predict what the range of outputs can be. The next steps in development will most likely be to improve classification of freetext and use it more extensively within the app. We've also considered having individual nodes which use generative output (at least in part for the output), to go more to the generative end of the spectrum, but retain control of the conversation flow, however this is still at the concept stage for the moment.

We've seen some graph based / smart script engines where you can enter subgraphs where the output is essentially seq2seq until various trigger conditions are met (e.g. the text is classified a certain way), before continuing to another part of the graph, and this is another way the two approaches can be blended. It's not something we're looking at at present, and can also be potentially quite unpredictable.

[tdaltonc]

"Generate chit-chat until the intent classifier gets what you need" is an interesting idea. Should work especially well in therapy where asking the same question for different angles is pretty well tolerated.