Don’t Experiment on your Customers with Machine Learning

Conversational AI stage 3

This article is part of a series that explores the main steps to starting a conversational AI strategy. Find the complete blog series here:

In the third part of our “Getting Started in Conversational AI” series we consider the complex issue of machine learning. We look at the problems surrounding it, the role that some development tools expect your customers to play, and why a pure machine learning approach isn’t necessarily the right choice for many enterprises.

Of course, machine learning has opened up a world of possibilities – particularly in areas where large sums of data need to be mined and analyzed. Not a day goes by without an article that mentions how machine learning has enabled some amazing feat to be achieved.

And it’s all true.

For instance, speech recognition – the technology for mapping acoustic sound waves to text characters and the first step in processing speech – successfully uses machine learning to learn from the vast amounts of voice inputs and improve levels of recognition.

So machine learning most definitely has its place.

Except that’s not the whole truth. When it comes to conversational AI, there are issues and limitations that enterprises need to be aware of if they are to develop conversational interfaces successfully. And most of them revolve around data, or rather the lack of it.

The challenges of machine learning

One of the key problems is that the only language a machine understands is binary. It has to be taught how humans speak. But machine learning algorithms can only discover the many different ways a human might ask a question if they can learn from copious amounts of examples. This training data needs to be classified and cleaned, which can also be a long, laborious process performed, ironically, by humans.

The challenges of language diversity become more apparent in real-world situations. It could be your very first day on the helpdesk for a clothing firm and with no training whatsoever you’d still understand that “Where’s my delivery? “or “What’s happened to my parcel?” and “Is my shirt coming today? “amount to the same problem.

In fact, you probably wouldn’t notice the multitude of ways the same question is asked. Even misspellings wouldn’t faze you if it was a live chat environment.

But for a machine, this level of complexity can quickly cause problems, unless it has been taught. And not only that, it needs to learn your ethics, ethos and brand. Just as you were guided on life’s rules and values as a child.

Remember that kid. The one your parents would never let you play with, because their folks just let them run wild. Well without control, that’s just what a machine will do. Maybe not in a 2001 Space Odyssey way, or at least not yet, but enough to make your brand famous for the wrong reasons!

The need for data to train machine learning systems has created a dilemma for many enterprises. How do you collect conversational data without a conversational system?

Experimenting on your customers is not the answer

Few enterprises have available the vast resources required to curate and train the data. Even then, it takes time, a precious commodity that most businesses can’t afford in a competitive marketplace.

So they make a fatal mistake.

They develop an app and expect their customers to train it for them.

Typically, the interface will only be able to recognize questions asked in a specific way, anything else will be an anomaly thrown out to a generic ‘safety-net’ answer or the customer immediately transferred to a live agent. These errors are then flagged to a human to correct.

Depending on the development tool used, it may mean additional work adjusting the system by a specialist linguist technician or hours of work coding in the different ways a customer might ask just one question. All of this takes additional resources and time. Meanwhile customers are still using the app and receiving the same unsatisfactory experience.

Or the app may answer the question, but it won’t necessarily be the one your customer asked. In which case, unless the customer helpfully told you “No this didn’t answer my question”, you may never know it’s answering with the wrong response.

Either way, though the app might learn a little more with each interaction, at what cost is this to the business?

Don’t get me wrong. We all experiment on our customers. Searching for the little things that make a big difference to the bottom line. It might be a rejig of the website, or perhaps a new internal process. But these are typically minor changes. In the big scheme of things, if something doesn’t work out, it’s no deal breaker.

But launching a conversational app that doesn’t have the conversational skills to understand your customer’s questions, let alone answer them, is a recipe for disaster. Particularly when speed and convenience are winning over brand loyalty. Customers will stop returning if the experience is less than satisfactory. And worse? They’ll tell their friends too.

Solving the training data conundrum

At Artificial Solutions we overcome the lack of training data in several ways. We have our own language resources for you to use. These have been built up over the last decade from billions of conversations. When you’re developing dialogue flows in our conversational AI platform, Teneo, you simply need to give a few samples as how a question might be asked. Teneo uses the language resources to work out all the many different ways that someone may phrase the question. It can do this in over 35 languages.

These same language resources also allow your conversational interface to be able to handle the erroneous questions that customers sometimes ask chatbots such as “what’s the weather like”, their favorite color or even out on a date. All of the responses can be customized allowing you to put your own personality stamp on the application.

Teneo doesn’t rely on machine learning alone. It’s too much of a wildcard for enterprises that need control, as Microsoft’s Tay proved a few years ago with its racist chat on Twitter. Instead we take a hybrid approach, combining linguistic and machine learning at a native level. The machine learning allows the application to make smart and complex inferences, while the rules ensure the system maintains a consistent and correct personality.

Our drag and drop style graphical interface allows developers and business users to collaborate and easily construct the intelligent architecture behind human-machine conversation to ensure the right response every time – with or without the data.

While Teneo does use real-life customer interactions to improve the conversational AI application, thanks to our patented hybrid approach, it’s only to optimize the system, not to train it.

Machine learning delivers significant benefits in conversational AI, but its fallibility when used alone, should give enterprises pause for concern. By taking a hybrid approach, combined with built in language resources, enterprises can not only have their conversational AI application up and running faster, but remove the risk of alienating their customers in the process.

Continue the blog series here:

Machine Learning

Conversational AI Center of Excellence

Begin Your Conversational AI Journey

Master the key steps towards creating the ultimate conversational AI solution for your business.

Register Now