/ How do you create question variations for an AI chatbot?
Alexander Wijninga
|
3 min
You are all set: building a chatbot! That virtual colleague could save you a great deal of work, so you cannot wait to start. Before you get going with your chatbot, though, it is important to map out your customer questions and draw up good question variations. In this blog we explain how to do that!
The name says it: a question variation is a variant of a particular customer question. The question is phrased differently but has the same intent. Let us give you an example. The main question might be: “what are your opening hours?”. A variation on it is: “at what times are you open?”. The question is phrased in another way, but the customer means the same thing.
Our AI chatbot works with a Bag-of-Words model. This model looks at the number of times a word appears in a text. We will not bore you further with the technical aspects of this model, because that would make this blog unnecessarily complicated. What it effectively comes down to is this: the higher the number of different words your chatbot has in its “brain”, the better the chatbot can answer customer questions. So when the chatbot knows more question variations, it will also understand more combinations of words and therefore become smarter.
First of all it is important to gain insight into your customers' frequently asked questions. You may already have an FAQ on your website, but you may also still have to draw one up. In the latter case: collect as many questions as possible that have been asked by email, WhatsApp, chat and telephone. Note all these questions in a single file. Involve colleagues and customer service staff in this process too: which questions do they see coming in?
Now try to cluster the frequently asked questions by subject. You probably give the same answer to “what is the delivery time?” as to “how long does delivery take?”. You cluster these questions under the label delivery, for example. That gives you a list of frequently asked question subjects.
To create good question variations, you will have to use various synonyms. Let us look at the question: “how do I send an image?”. Not every customer will use the word image. To gain more insight into what customers might ask, we open a thesaurus and go looking for synonyms for image. Synonyms for image are, for example: photo, illustration or attachment. Do the same for the word send, submit, add, and so on.
Besides using synonyms, it is also important to map out the conjugations of the verb. A customer might ask: “how do I send an image?”, but it may also be that the customer deliberately keeps the search short, for example: “sending an image”. You want the chatbot to recognise both verb forms.
In the previous two paragraphs, various words, verbs and synonyms have passed by. Once you have noted the variations down for yourself, try to write out as many ways as possible in which the customer might ask a question. Keep the conjugations of verbs and the synonyms in mind here too. Try not to use the same words too often and draw up at least 6 question variations.
Example:
“How do I send an image?”
“Can I send a photo?”
“Is it possible to send an attachment?”
“Where do I add an illustration?”
“May I add a media file?”
“Where can I add photos?”
Once you have added the list of question variations to the chatbot, the testing can begin! Have as many (independent) people as possible test the chatbot. Think of colleagues from other departments, but also family members or friends. Everyone has their own way of asking questions and you want to map as many of them as possible. The questions the chatbot does not recognise (and therefore cannot match to an answer) will appear in the Watermelon platform under “mismatches”. Add these to the question variations and the chatbot is ready to go live!
Would you like to get started building an AI chatbot? Try it yourself and create a free account. Would you rather read on about successful chatbot examples first? Then request a whitepaper!
Share this article
