Understanding language is hard for a computer. Understanding the language of IT Support is an order of magnitude harder.

The rules of syntax and grammar are always changing and words have many meanings that can be combined in infinite ways. Of course, what someone says is only half as important as how they say it. Context, sarcasm, and idiomatic speech can further complicate matters by adding ambiguity. Now imagine all of this, but include the difference between MFA and 2FA, between Zoom and Teams, between an incident and a ticket. This is the maze of complexity that a world-class IT support agent expertly navigates as she picks up the phone to help triage issues.

For a conversational AI system to do the same, it must tackle the nuances of language and the language of IT. This is both the great challenge and promise of developing deep learning systems that provide Natural Language Understanding (NLU) for ITSM.

The words in a support request don’t tell us everything

A good NLU System will use contextual cues to figure out what ambiguous words mean.

We don’t always say exactly what we mean, because we expect the person on the other end to fill-in-the-blanks. People do this instinctively by relying on contextual clues. This might include things like a person’s background, what their job title is, and even whether it’s before or after lunch.

If an IT support professional has to triage a user-generated ticket, she will have to understand how many people are affected by an issue, if it was submitted in the past, and what department(s) might be involved. This would be incredibly difficult for a conversational AI system that relies purely on the language of that ticket.

Counterintuitively, by training NLU models to learn from contextual cues that have nothing to do with language, we get closer to how humans intuitively extract so much from so little.

Resolving issues needs real world knowledge

A good NLU system will detect that “her” doesn’t refer to Tamara.

People don’t like to repeat themselves. “My laptop is slow because my laptop has no more memory in my laptop” is an awful sentence to say and hear. Words like, “it”, “she”, “his”, or “these” help us escape this nightmare. Since we live and breathe language daily, humans find it relatively easy to resolve these types of references. It also helps that we know how objects exist and behave in the real world, and can expect the person across from us to have that same understanding. This is not so easy for computers.

If you say that you spilled coffee on your laptop because it was too hot, we’d like for a sufficiently intelligent NLU system to understand that the coffee was hot and not that you have an overheating computer.

Advanced deep learning models equipped with reasonable world knowledge can help us build conversational AI systems that have a nuanced understanding of language.

One support request can have many sub-requests

A good NLU System will understand all distinct user requests.

When we speak, we often convey more than one thought in a sentence. This makes it easy to stitch together complex ideas and goals. Humans are capable of quickly figuring out that when you say you are going to the bank and then going shopping, that these are two distinct actions. We don’t ignore one or the other, nor do we get confused that you might be shopping at the bank.

While system-generated tickets are usually straightforward, user-generated tickets are complex and can sometimes include multiple requests at once. A trained ITSM agent doesn’t just fixate on the first keyword she comes across, but will attempt to resolve every issue. Drawing a fine line between requests in this way is incredibly easy for a human, but hard for a computer.

By running multiple machine learning models in parallel and averaging their predictions, an NLU system can start to detect when a group of sentences contains multiple concepts at once.

Employee issues are usually vague

A good NLU System will resolve amibuity by narrowing down the right answer conversationally.

One word can have many meanings. Case in point, when the next Oxford English Dictionary is published in 2037, the word “run” is anticipated to have 645 distinct definitions. Even with all this expressive power however, in the IT context, employees usually only provide a vague description of an issue they might be having. This forces ITSM staff to go through a lengthy back and forth over email and many days before effectively resolving a ticket.

Detecting the ambiguity in language and then engaging a person in dialog to get to the bottom of their issue is an incredibly challenging task for conversational AI. However, by learning from conversations, NLU systems can detect patterns in ambiguous requests to help narrow down the space of solutions.

Employees don’t always say what they actually need

A good NLU System will infer what a user needs from what they say.

The reason why humans are so effective at communicating with one another is not because we use precise language. In fact, as we’ve seen already, that is very much not the case. We communicate rapidly because we are extremely good at predicting what a person will say next in any given moment. Years and years of evolution and generally being around other people has strongly primed us to intelligently jump to the right conclusions.

IT agents constantly infer what users actually want from the language they submit in incident or service request tickets, and this fundamentally makes natural language for IT a hard problem to solve.

Deep learning models which learn complex patterns between what is stated and what is implied in IT support tickets help us create NLU systems that can solve problems quickly and efficiently.

We’ve now seen why NLU for IT Support can be so difficult, as well as how to think about solving this problem at a high level. Next, we’ll look at some breakthroughs in deep learning and language modeling that help us build an NLU system for ITSM.