Business vs linguistic approach

How to build NLU effectively? Can I just start with copying my IVR?

Often, when building a bot, the SentiOne client is guided by the current IVR helpline structure. In practice, this means that when building a bot, it is guided by the topic or product discussed in the conversation.

For example, if we have two customer service agents, Mark and Tom, and

• Mark is responsible for type X insurance,
• and Tom is responsible for type Y insurance,

... then the intents are built on the type of insurance, i.e. in such a way that the phrases regarding reporting damage in X insurance will be grouped in the intent of X, and the phrases regarding reporting a loss in Y insurance will be sent to Mark through the intent of Y.

Unfortunately, this causes conflicts between the phrases, the model confuses them and may direct Mark's topics to Tom and vice versa.

Why is this happening? The phrases in both cases are very similar to each other, e.g.

• I want to report damage under the travel policy.
• I want to report damage under my health policy.

The model sees that these two sentences differ only by 1-2 words (our OBJECT), and therefore considers them similar to each other. And rightly so! :)

👍

How to solve this?

Phrases of this type should be grouped within the same intent, because they are similar to each other, express the same communication purpose, and differ only in the type of product (policy type) against which the activity is performed.

How to avoid those problems? Tips!

In order not to fall into the trap of designing the helpline in a manner similar to building an IVR, i.e. "business oriented manner", it is recommended to:

1. Begin building a set by defining the basic intents (answer the question "what will someone want to do on a helpline using a bot?" And write out a list of these intents),
2. For each of these intents, list at least 3-5 ways of expressing the intent (constructions),
3. Write about 5 phrases for each of the constructions
4. Determine which products will be related to individual intents and group phrases in terms of actions (e.g. cancel, notification of damage), i.e. an action performed on a particular object, and not based on objects.
5. The set of phrases prepared in this way should be submitted to the team of bot designers and linguists on the SentiOne side for verification.