Automatically extract Dialogflow intents from chat logs

In this article I provide some ideas on how you can automatically extract Dialogflow intentsBoth Dialogflow ES and Dialogflow CX have the concept of int... More from chat logs.

An example dialog dataset

I took this example dataset where people called in to customer support at a telecommunications company (the dataset already has redacted personal and other identifying information). This is actually a very good dataset since it a) is obviously realistic and b) provides a lot of insight into this topic on the whole, as you will see later.

Here is a sample list of sentences from the dataset.

Results from using the tool

So once you upload this dataset to the Autotrain tool, you will get a CSV file which is an extension of the 4 column CSV format which you can use to define and organize your Dialogflow intents.

Here is a screenshot of the resulting CSV file based on the list of messages above (I used a total of 500 messages). Note that the tool is able to group multiple phrases based on their intents (e.g. pay bill) and also outputs the source identifier (SentID) from the original document for each row.

What you can infer from the output of Autotrain

We can make some interesting observations based on the output file. These observations will in fact make it easier to clean up the output and improve your Dialogflow bot’s training process.

The user message can contain multiple sentences

The conventional approach is to specify a single sentence in a single Dialogflow training phrase. As you can see, that isn’t ideal.

A single sentence can contain multiple topics

This is even more challenging, because splitting up the user input into sentences isn’t always going to be sufficient

A lot of user messages span multiple intents

If you took the output produced by the Autotrain tool, and grouped the results by sentence ID, you will notice that a single user message spans multiple intents

The multiple intents may need multiple handlers

That is, if you do have multiple intents inside a single user message, you might need to process them as separate intents because they might require completely unique actions. For example, take a look at this user message:

To handle this user’s message, we have to understand two separate intents. 1 The customer has paid their past due amount, and hence 2. The service needs to be restored.

Contrast that to this other user message.

While the second message is also talking about restoring service, it is due to an entirely different reason (outage).

How to create your Dialogflow agent from the output CSV

As I mentioned earlier, the format used in the output CSV file is based on the 4 Column CSV file used in the FAQ Bot generator tool.

So you need to delete everything except the first 4 columns, and you will have the 4 column CSV file you need for generating the Dialogflow agent ZIP file with a single click. You can of course modify the CSV file and add a response for each intent etc. before you convert it to a Dialogflow agent ZIP file.

You might have noticed that the number of phrases which are actually categorized into non-unique intents (that is, have the same pattern as another phrase) is only about 20% of the dataset. In the follow up article, I will explain how you can cluster more of your phrases into existing intents.

About this website

I created this website to provide training and tools for non-programmers who are building Dialogflow chatbots. 

I have now changed my focus to Vertex AI Search, which I think is a natural evolution from chatbots. 

Note 

BotFlo was previously called MiningBusinessData. That is why you see that watermark in many of my previous videos.