Train the bot

By going into the main menu and to ‘’Conversations’’ we can train the bot by reading the dialogue entirely.

We can choose which period we see by changing "From date" – "To Date" and then clicking "Get conversations". If we train once a day about 10 minutes, which is quite normal we don't have to change the period. The default is that it shows yesterday.

"Get bot transfers" shows the dialogues where the customer has chosen to be connected to an agent.

"Active dialogs" shows ongoing dialogues live.

"Low confidence" shows dialogues where the bot has not been able to answer the question at all but answered by presenting an error message ("I did not understand what you meant...")

This is how a query with low confidence looks.

We can see that the bot interpreted the question with 61% confidence level that it is about adult education. But for the bot to respond, in this bot, 65% level is required.

To train the query click the arrow on the right side of the dialog box. Then you get a drop-down menu where you can choose the right workflow. In this case it would have been "Vuxenutbildning-vilkaKurser"(AdultEducation-WhatCourses)

When you have chosen the right workflow, a blue button appears on the right "SAVE UPDATED MODEL". This is done after every change when we train in "Conversations".

The "GET LAST QUERIES" button brings up the latest X number of free text queries. You can set how many questions you want to see yourself. You can select the next page in the bottom right corner of the picture.

In the table, you get a row for each free text query in the order of which they were prompted to the bot. You can see which intent the query matched with and with what confidence level. If it is low confidence level but the correct answer, you can click on the question which then prompts you a window where you choose the right intent and click save to train on that sentence. Unfortunately, you need to click into Query Log again after this moment.

Once you've built all your intents, workflows, and put in the training data you think is enough to activate your new intents in the bot, you'll need to load your changes into the NLU model. You do that with the “walk” button.

This allows the bot to learn what you've changed. It takes between 5-10 minutes to train the bottom. Therefore, this should be done sparingly. But of course, it needs to be done a little more often when you build in a completely new topic.

When the training process is done you can try your new intents in "Visual Trainer"

The visual trainer is the most effective way to test out your new workflows. After you have ran the training process, you will be able to query the bot in visual trainer. A good idea is to let other team members of your team test the bot in visual trainer.

When prompting the bot in the visual trainer you will get a value for the confidence level. This value represents how well the bot understood your question. If the bot fails to understand it will prompt you with the fallback intent.

If the answer you get from your bot is weak even though you think that the question you asked should be answered with a high confidence level, select the drop-down menu and select the correct intent. If your bot responded with the right intent but still shows a weak confidence level, you must choose the right intent again to improve this intent. Once you have selected your intent in the drop-down menu, the question you asked in the visual trainer is saved as new training data in your intent. When you have finished testing your bot in visual trainer, you need to train your NLU model by going to intents in the main menu and clicking on "train your model" icon.