What happens after you build a chatbot for a client? I mean, when the architecture is set up, the NLU module has been built, the conversation flows are designed, and everything is ready. All is left is going into production, where you will have thousands of users' interactions that will train your bot and make it super smart, right?
Right - if you are prepared to go through hell.
Many bot builders fear day 1 of production and they are not very wrong to do so: the bot doesn't understand most of the queries so there are bad experiences and complaints, angry clients that don't always understand the inner workings of AI systems, users who decide not to ever talk to the bot again, exhausted teams overloaded of work...
The common analysis for this situation is: chatbots are so trendy, they generate expectations we can't satisfy. We have to be able to communicate better, and make everyone understand chatbots need some time to learn, and one can't expect them to be great from day 1.
That's like giving up.
If there wasn't anything in the market that would come as a solution for this situation, we would agree; but there is, and it's a rather simple one.
To solve these kind of issues you need to be in charge of what the bot can do. The problem is: everyday a chatbot is launched, its designers have no idea of what the bot understands or not. To learn that, you would have to try it. Every time.
If it can handle simple orders today, there's no guarantee it can take those same orders tomorrow; or if it understands one word, that doesn't mean it will behave the same way with a synonym. This happens because its training is always running and may have changed something in its behavior.
On top of that, at first, you think so many people are going to interact with your chatbot in just a few days that it will achieve a super high level of understanding almost immediately. Let me tell you: that's simply not gonna happen. The few users reaching to your bot will run away bored at the very first "Sorry, I can't understand you".
So what's this simple solution we are talking about?
The problem are not (or not entirely) the high expectations. The problem is the NLU module. That's what's causing you lose control over your bot if it's entirely built on Machine or Deep Learning, and nothing else.
Luckily, you can forget about losing control over your AI system if you use a NLP middleware to improve your NLU module. Think about it: you use other NLP solutions as APIs in some specific point of your pipeline, so why not do the same with AI?
At Bitext, we have been able to design a query rewriting technology that transforms any kind of sentence into a simplified version of it, the kind of query the bot can understand, in a convenient structure, adding the elements humans usually don't mention and removing the ambiguities only we as humans know how to handle (until now, of course).
So stop living in fear. You and your team don't deserve it, not to mention your clients or your users. Do you want them to trust you from the beginning? You know what you have to do!
We are not quitters. We want to move forward in the quest for natural language interaction. And we're not crazy; we're perfectly aware that these are just baby steps, but we won't stop until true natural interaction is finally achieved.
We have already implemented this technology in a pre-existing bot, making it fully conversational and capable of correctly handling common issues such as negation and double intents. Learn how by downloading the use case below!