Let's take a glimpse under the hood of this customer service automation engine.
Orion is built on natural language processing (NLP) technology. NLP is a subfield of artificial intelligence. NLP touches upon two umbrellas of techniques namely machine learning (ML) and deep learning (DL).
Every day, new innovation is happening in the AI domain. Since the launch of the open-source machine learning library Tensorflow in 2015 new libraries, platforms and techniques developed by a variety of research centers, universities and organisations are popping up every 3 months.
It is our vision to master the various state-of-the-art technologies out there, in order to bring a best-of-breed mix to the table. This allows us to stand on the shoulder of giants and deliver the best possible results to power customer service needs.
Staying up-to-date on latest developments and putting together the right mix of technologies and techniques is our competency. Combining computer science with our experience in the customer service domain together with the ability to integrate, is what makes Orion today.
The basis of customer service automation is the ability to automatically recognise customer questions. Everything else follows from that.
Under the hood we build 1 model per language that can handle a variety of digital text channels.
Using our proprietary technology for e-mail we can extract the essential paragraph from an e-mail thread.
Webforms and contactforms are among the cleanest forms of customer question input.
Because of the short sentences used in chat and social media, the essence of the question is quickly captured.
You might not be aware, but e-mail is a particularly hard nut to crack when it comes to natural language processing.
If you just input any kind of text into an AI model the result is in 99% of cases useless. "Garbage in, garbage out" as we sometimes say. Reason is that such models do not know what parts of text to focus on. This we want to avoid.
Since the style and structure of an e-mail question differs from a contactform, which again varies from a chat message, our first step is to transform that. This is what we call pre-processing the input.
The pre-processing allows us to make abstraction of the original channel and feed the processed question input into the single recognition model we have custom built for each of our clients. The recognition approaches the state of the art.