This is part 2 of the series, and if you haven’t read part 1- How to start right, we recommend checking it out first here.
When we talk about chatbots, intent recognition is where the magic happens. Any decent chatbot will allow users to type their questions in natural language, understand what they want (intent), and provide the necessary answer/support. Today we will discuss the potential problems organisations run into with this intent engine and the way they are set up.
The intent engine is a black box in most DIY platforms, and you have very little control over it. Most customers don’t realise they have this problem until they are into their 5th or 6th iterations post-go-live, making this a hard one to spot. Typical issues faced are,
- Time-consuming to create intents
- Hard to identify conflicting intents
- Managing conflicting intents is a nightmare
Here is a detailed explanation of the problem for those in the early stages of the journey and yet to encounter this issue. Most DIY platforms give some form of out-of-the-box pre-packaged intents or models to jump-start the implementation. This is a good thing as your bot can understand basic conversations from day 1. When you start configuring intents to make it your bot, i.e. understanding your customer’s question and addressing your use cases, you need to add new intents and strengthen existing ones. To do this, you are expected to type in training phrases “manually“. This becomes a time-consuming task, and you have no visibility or control over how the system recognises the intents, especially when you have two very similar intents. For example, in a travel domain, you may have different processes when someone is about to miss a flight vs someone who has already missed a flight (Missed flight intent vs About to miss a flight intent). In the first scenario, the customer still has time to change their ticket; in the latter, you will have to discuss refunds based on their ticket type. For the first one, the customer might say, ” I am going to miss my flight “, and for the second one, they might say, “I have missed my flight “. There is a high likelihood of the system triggering a wrong intent unless you have a high level of precision to differentiate these.
When project teams hit these limits, their only option is to raise a support ticket with the vendor. You can expect the most common response is either “you have too little training phrase” or “you have entered too many training phrases” for an intent. The team will have to go through a tedious trial and error process to tweak the model, publish and test.
The problem arises because there is no early discovery/analysis of the conversations and most tools offer only a single way to create the intents, i.e you manually create an intent and enter the training phrases.
As you start to scale by addressing new automation areas, solving new customer problems and making the bot synonymous with your brand, handling conversation subtleties is very important. A better approach is to use a Transparent Machine Learning tool that you have full control over, like Verint’s Intent Manager. The whole process starts with your customer’s conversation data. You feed in the customer data and any metadata you may have into the tool to automatically categorise and discover intents (we covered how to choose the proper use case in part 1).
At this stage, you have several model techniques to understand the conversations better and create intent
- The multi-word cluster that groups conversations based on system generated ideas
- Single-word frequency model that helps you identify commonalities among data using term occurrence in relation to one another
- Pattern model – you use the conversation to define patterns, and the system uses these patterns for ongoing recognition.
- Classifier model – system predicts the best intent by comparing customer conversations to current intents.
All these models help you analyse and segment several thousands of conversations quickly & easily. If you want to use our platform to create the intents, all you have to do is drag and drop the conversations to the right intent. You can even group the intents in a logical hierarchy to make it easy to manage them as you scale. When you generate the intent model, the system highlights conflicting intents that could potentially affect your model. Within the same interface, you have the option to test the model using sample conversations. During these tests, the system highlights all the intents triggered and orders them based on weightage which helps you adjust your model to define precise outputs based on the scenarios you wish to address before publishing the model.
The advantages of this approach are:
- You have a better understanding of conversations and patterns before you start creating the intents.
- Have multiple modelling strategies to better understand the customer data and mine for information.
- The process used to analyse the conversations can be used to create new intents.
- The intent is created using your actual customer data – this will have higher accuracy and better coverage.
- Quicker to develop new intents when compared to manual entry of intents and training phrases.
- Faster to manage ongoing changes as you can drag and drop conversations into new intents when you want to modify or create new intents.
- There are no limits on the training phrases you have under an intent.
- There are no limits on the number of intents in the platform.
- Predict potential conflicts before deploying the model.
- Test the model to get a better understanding of the intents that are being triggered.
- Most importantly, a transparent machine learning model that you can control.
If you would like a see the Verint Intent Manager in action, click here .If you are using other chatbot technology and are facing issues with scaling, we can still use our AI tools to help you innovate better in your current platform – reach out to us .
If you have already invested in a chatbot platform but facing challenges with scaling, we can reuse your current investments and migrate to a scalable and transparent Verint IVA platform – reach out to us .
– Arvindh Janarthanan
Connect with me on LinkedIn