Context over verbatim: we want to understand the customer

Paul Stuefer: Context over verbatim

An interview with Tinka's father

Paul Stuefer is a Product Manager and is in charge of the Tinka project. After training as a chef, waiter, and media IT specialist, he went on to study graphic design – only to discover that he wanted to have more to do with technology. So, fatherhood with T-Mobile Austria was right up his street.

I have just asked Tinka: Who is Paul Stuefer? Her reply was, “If you ever meet the all-knowing Paul without a headset and if he’s not on an E-business scooter, then you should whisper hallelujah quietly to yourself and grab him for an answer to your clever questions. #ask Paul.” How does it feel for Tinka to know everything about you?

That is a nice gesture. Our interns fed Tinka with this information a while back. I think, in future, things should also be approached at this kind of level. This way we can speak to our customers more personally.

What exactly are your duties at T-Mobile Austria?

I’m pretty much Tinka’s father and I am the primary project manager for the entire project surrounding her. There is an entire ecosystem, which is more or less like a library that Tinka can access to pull out the right book.
In turn, this ecosystem uses the FAQs that we have newly put in place. We have redesigned the T-Mobile customer forum from the top down, enhancing it with many functions. We have also created a knowledge cluster. Tinka access this; thereby receiving a variety of information that helps her to answer customer enquiries. Tinka does not have access to sensitive customer data.

How exactly would you describe fatherhood for you?

It really is live having a child. I am busy all day long. I am always attending to little bumps and bruises. There are plenty of things that I can still teach our little Tinka to do. None of this would be possible without our great team – including our colleagues in Germany. Without their help, the Tinka we started out with two years ago would never have come this far. Since then, we’ve taken Tinka out of kindergarten, and now she can go to school. This is a great feeling and it is really wonderful to experience how this system, our little Tinka, is growing and developing.

Her role is natural language processing. So how does she talk? What does she understand?

Tinka has one or two Austrian quirks, and her accent has a hint of an Austrian flavour. However, from a purely technical perspective, we are also trying to get this in the background. This is why we train the natural language understanding components using Austrians, and Austrian conversation data. Only this way can Tinka understand our customers – even if they are from the regions around Vorarlberg, Tirol or Vienna and their wording and grammar vary a little from standard German.

Which methods or models do you use to understand language?

We use an NLU component that is intention-based. This means we try to understand the customer’s intention, then clarify it using counter-questions to work out an optimum answer. It is currently still possible to select these manually by clicking on answers; however we will increasingly vocalise these in future and train them using annotations in the NLU component.

What does manual selection mean? Will a possible answer be entered for every possible question?

No. We have defined key words until now. When one of those keywords is used, a pre-defined answer is played. We are currently in the process of changing this. In future, we want to make use of the knowledge in its entirety. Tinka’s brain will be transferred to an artificial intelligence component, to classify the customer’s intentions. Take for example, a customer writing that his or her bill too high. In such cases, we have manually defined a dialog. If the keyword “bill” is used, Tinka asks a counter-question for clarification and then enters into a dialogue. We take this as a basis for starting out, and we re-classify it. The customer might also ask, “Why have I paid too much?” He basically means the same thing; however, he doesn’t mention the word “bill”. In future, we will harness website content, forum contributions or similar, to make it a more dynamic and autonomous process. In other words, if we don’t know something, Tinka can access the T-Mobile customer forum, for example, search for an answer there, and provide this to the customer. Our main goal will be not to enter everything manually and to define exactly how the answer should be formulated. This is why we have decided to use artificial intelligence to understand the entire sentence and not just individual words.
She should also be able to draw knowledge from further sources of information at her disposal. For example, what the client’s customer journey looks like. Perhaps they might have previously researched a mobile phone at the T-Mobile shop, and they have switched several times between iPhones and Android. If they then ask Tinka which phone she recommends, she should have this information and be able to explain the respective advantages of each phone. By using her knowledge of the customer’s telephone and surfing behaviour, she should also be able to recommend an appropriate contract tariff. This means that in future Tinka really could take on the role of an assistant. As a digital concierge, she will be on hand to assist the customer at any moment.

Does this mean that Tinka’s understanding is not verbatim; rather, she understands the context?

She understands my question, but she doesn’t understand every word in the sentence. This is sufficient to understand exactly what the customer wants. It forms a basis for the answers or solutions. However, this approach fails, Tinka calls upon human support. The dialogue is transferred on to human colleagues in our call centres or to live chat. Even after this handover, the conversation with Tinka is retained, thereby avoiding the need to restart the consultation from scratch.

What role will emotion recognition play in this respect?

Emotional analysis will be one of the issues we address in future. If customers ask about bills, their tone of voice is decisive in how the question is meant. Perhaps they would like to know how much their bill is, in which case Tinka can help; or perhaps they want to make a complaint, in which case, the conversation is passed over to a human colleague faster, so the customer receives better support.

How does Tinka learn to speak in a manner that fits with her character?

We train the system and carefully look at the information we give her. This starts out from a very narrow framework. We examine logged conversations and define the underlying intention behind them, then we teach the system what they mean and how it should react. This can be scaled up. Then there is a dialogue module, which controls speech modulation. It is here where Tinka’s character is captured. In other words, there will always be a certain essence of Tinka in what she says.

Does this mean that she also learns from various dialogues?

Exactly. Above all, she learns what we give her. If I give a child medieval texts to read, then he or she will probably start greeting me as “my liege”; however, If I give the same child contemporary literature, the language learned will also be more modern.

How high are the traffic volumes?

Tinka currently chats 24/7, if customers ask more than one question, this means that they enter into a dialogue. We also have a lot of customers who prefer to just click their way around and don’t need any support. We are looking towards investing more in quality than quantity. In particular, we want to understand our customers better.

How important do you think it is that Deutsche Telekom and T-Mobile Austria are moving very quickly in this area – practically in the vanguard?

More than two years ago, we had already launched Tinka to the market – sooner even than many people had even started talking about chatbots. The last six months have seen a flood of chatbots, bot messengers, and the like. All of the established messaging platforms have now launches a bot. We can be confident and hold our own during this phase. Ultimately, we already have a year-and-a-half of experience, and we can run the bot framework on various channels. This means that Tinka has now recently also started helping customers via Facebook Messenger.
You have to react extremely quickly and flexibly to the market. If you don’t, a company doesn’t stand a chance. Deutsche Telekom very quickly recognised the potential on offer here in Austria. It set up this project in order to progress quickly and create added value. This is the correct approach, as is the investment in artificial intelligence. We took the decision to use AI early on – long before it became a craze. We will quickly bring to the market solutions that capture people’s attention because they are state-of-the-art or even go a step beyond.

Add a comment

your browser is not up to date
to enjoy this website you will need to install a modern browser.
we recommend to update your browser and to install the latest version.

iOS users, please male sure you're running at least iOS 9.

Mozilla Firefox Google Chrome Microsoft Edge Internet Explorer