Using Artificial Intelligence in LiveChat: Smart Tag Suggestions

By in LiveChat Blog > Live chat use cases,
Artificial intelligence in LiveChat

Artificial Intelligence (or AI for short) has become quite the buzzword recently. Everyone in the tech world seems to want to use it. A ton of different applications cropped up in the last couple of weeks, every more amazing than the last.

We’ve seen AI that can tell us what it sees on an image, one that can suggest an answer based on an email contents and one that could tell us the weather if we ask the right questions.

We wanted to check if there is any way we could use artificial intelligence in LiveChat to improve it in any way. As it turns out, there’s quite a few applications that can make the work of agents using LiveChat much easier and efficient.

A couple words about AI

The kind of AI we’re seeing cropping up in various software and IT projects is quite different from the one we know from sci-fi movies.

It’s still very far from a real AI that would be able to grasp complex concepts. However, it has the ability to learn. What you could do with something like that? For example, you can teach a machine to recognize specific patterns or words in a message and suggest a relevant response. Using that, you could have an AI that would be able to tell you what the weather is like or what is the best route to a specific place.

How does an AI learn? You first need to feed a ton of information into it. For example, thousands of weather-related questions and answers. After a while, the AI should be able to recognize which answers fit which questions.

You could say that the same thing can be done by pre-programming the answers and assigning them to specific questions. And you would be right. However, an AI that has been fed enough data can accurately answer questions that it has never seen before.

Using artificial intelligence in LiveChat

We wanted to try out this learning ability of the commercial AIs in LiveChat. Grzegorz Wyszyński, one of our Software Developers, came up with an idea to use AI to offer smart tag suggestions during chats. Here’s a basic idea behind how it works.

Tagging chats lets you to categorize the conversations you had with customers to get a better understanding on what’s happening on your LiveChat and to be able to easily find them later on. However, when having several conversations at a time, it can be hard to pick the right tag for every chat.

To help with that, we wanted our AI to suggest tags automatically based on previous chat history. This way, whenever a sales chat comes up, LiveChat would suggest a ‘sales’ tag.

It works similar to the way Google Inbox suggest short replies to emails automatically.

The AI learns how to tell different types of chats by checking previous chat history. To offer suggestions, the AI needs to see at least 100 tagged chats in the archives. This is a fairly low threshold but this allows more customers to be able to use it immediately. The more chats you have tagged, the more accurate the suggestions will be. It works for both the default LiveChat tags as well as the ones you’ve created, no matter which language you use.

We tried a couple of different AI models to make the project work: neural nets, support vector machine (SVM) and logistic regression. We ended up using SVM because it provided the most accurate results and required little computing power.

Results and plans for development

As of right now, over 1000 businesses are using the automatic tag suggestion feature. It already provides some really great results as over 42 percent of all tags assigned by agents using LiveChat come from the automatic suggestions. This means that 7 percent of businesses using the feature are responsible for nearly half of all tags! It seems that automatic tag suggestion makes it much easier for agents to tag chats reliably.

The accuracy seems to be fairly high, even with licenses with a larger number of tags. Here’s how it looks in numbers:

Number of tags Accuracy
3 tags 72%
4 tags 66%
5 tags 60%
6 tags 55%
7 tags 54%
8 tags 53%
9 tags 50%
10 tags 54%

There are still some minor problems we need to iron out. For example, the AI seems to have some difficulties with a couple of languages, e.g. Hebrew. However, the future of the project looks very bright.

When it comes to our plans for the future of the project, we’d also like to set up something similar for canned responses. We’d like to have an AI that could automatically suggest canned responses to agents. However, that involves a whole another order of complexity. Moving from one-word tags to responses that include several paragraphs would require a lot more calculations and raw computing power.

To make that happen, we started working with a local University of Technology that has an AI department. They will use natural language processing (NLP) to determine if it’s feasible to use AI to suggest canned responses. And if it is, we will definitely want to include it in LiveChat.

If you want to try out the automatic tag suggestion yourself, you simply need to tag 100 chats. The option should appear automatically after a day or two after breaking the 100 tagged chats barrier. There are some rare cases where the system can’t provide accurate recommendations despite the 100 chats but we’ll do our best to make it more precise with future updates.

discussion

comments powered by Disqus