The fundamental change in German industry is taking place through service: because more and more...
Chatbots are almost as rudimentary as they were 50 years ago – the German newspaper ZEIT ran a headline recently. Do you remember Eliza, one of the first chatbots? In the mid-1960s, she served as a virtual psychotherapist, reacting to text inputs. Her answers usually missed the point. And today? Many experts consider bots to be the future of human-machine communication. In the last few years, there has been lots of hype about the use and role of artificial intelligence (AI). Bots are almost always associated with AI. But should they be? What can AI do for chatbots? What functions can it reasonable cover? And where are its limits? We will try to answer these questions in this article.
COVID-19 and the corona bot
Have you met Corina? Several municipalities and districts in Germany have provided a new corona bot service. And example can be found at https://www.landkreis-ludwigsburg.de. The system understands even vague questions about issues ranging from emergency child care to masks to test laboratories for symptoms to the current school situation and delivers appropriate answers in an intelligently controlled dialog. This allows it to provide reliable information of assured quality.
Corina can also answer questions about soccer. For instance, a person asks, “When will there be soccer again?” The unclear formulation prompts the system to respond, “You’re asking about ‘soccer’. Do you mean the Bundesliga, or would you like to play soccer outdoors yourself?” If the user clicks “Soccer outdoors” and enters a postal code, the bot provides detailed information, including the fact that the members of only two households are allowed to be together in public and that public recreational spaces are open again. Otherwise, there are no restrictions for outdoor activities – as long as participants keep a distance of 1.5 meters.
Upon being asked if a visit to the hairdresser’s is allowed, the chatbot responds not only with the answer (“Yes, if certain hygiene rules are followed”), but also provides an overview – in this case a list of specific rules.
And the bot can actually do even more: When asked whether a person can register a new car, the bot requests the postal code. After it is entered, the system reports the associated town and the currently limited opening times and asks when the user has time to visit. The answer is “Tomorrow morning at 11:00”. The machine translates this statement into a specific date, compares it to current appointment lists at the registry office, and confirms the appointment.
Even though this last element has not yet been implemented, it is no pipe dream, but entirely feasible. Thus, the chatbot serves as a dialog partner that can deliver specific practical assistance.
The reasons for this are obvious:
- the user doesn't need to know how to ask the question in order to receive the desired answer,
- the bot provides answers and does not simply find documents,
- the chatbot not only provides answers, but addresses the user’s situation and can also teach the user something,
- the system can help the provider determine what questions are asked,
- the bot can integrate any other function without the user noticing the difference.
But which of these added value elements that the bot offers are actually based on artificial intelligence?
Learning, learning, learning – the problems with AI
At the moment, there are two primary functions for AI-based bots.
The first can be described as classification or pattern recognition. Appropriate questions (“utterances”) are defined for each intent (that is, for each answer/solution).
An example answer would be as follows: “Insurance covers medical return travel not only for medical necessity, but whenever such transport is medically reasonable and justifiable. It also covers transport for an accompanying person if such transport is medically reasonable and justifiable.”
Associated questions include: Is medical return travel insured? What about medical return transport? Is return transport included? And so on.
The problem is that the system is not familiar with the individual questions but learns patterns from lots of data records and maps them onto the answers. For AI to have a positive effect, the machine needs a “critical mass”, or hundreds of data records (in this case, questions). In practice, the effort required for intent recognition is extremely high. Perhaps 20,000 utterances are needed for 100 answers – which means lots of time, which service teams usually don’t have.
Moreover, it is not easy for a bot to learn when it SHOULDN’T recognize an answer. For instance, “Visit grandma” and “Visit grandma in her retirement home” should receive completely different answers during the corona crisis.
To make things even more difficult, humans can scarcely control bot “learning”. This makes tracing changes time-consuming and risky. So questions like “How many people can ride in a car?” or “How many people can I go out with?” are almost the same, but the answers will be different. If a chatbot answers a question incorrectly, it takes a great deal of effort to correct the system so that it returns the right answer.
The second AI function in the example cited above, the one with the car registration appointment, is called “entity extraction”. Assigning a postal code to a certain town or region, or recognizing a specific date in colloquial terms such as tomorrow, the day after tomorrow, etc., can already be accomplished with AI and used effectively. But of course, there will be individual situations in which there are difficulties in assigning values, such as with the statement “I told you yesterday that I have time the day after tomorrow.”
Humans AND machines – the right answer once again?
When the term “AI” was coined in the 1950s, it referred to the imitation of human behavior. This means that AI describes not technology, but intelligent behavior patterns. In the past, even rule-based expert systems have worked according to this principle, or modelled decision trees. Bots also usually use rules – but because of vagueness in meaning, they are often associated with AI.
Let us consider rule-based systems: they do not learn by themselves but require human assistance. Interplay between bot and operator can improve the machine’s learning relatively quickly. The foundation is integration into a knowledge base that maps connotations, synonyms, and homonyms, makes them available to the bot, and allows easy content maintenance. Given its conversational interface, the bot learns from humans, but the reverse is also true: The machine lists the questions to which there are not yet answers and documents new terms so that the editor can assign new terms or add solutions that he had not yet considered in this context.
Thus, associating bots with AI does not go nearly far enough. The examples listed above show that AI is too inflexible to be used with multifaceted content that changes frequently. The advantages of chatbots are not due to AI, but to their usefulness in collecting specific questions, offering an information exchange interface, etc.
What is frequently called AI in this context can be more neutrally termed “data-driven services”. The services work only if sufficient data is processed, however. That is why these functions are used for relatively constant, general information. Data-driven services has proven effective in the automatic identification of related terms and synonyms. The utility comes from the interplay of analyzed data, human intelligence, and software – and that is the key.
The decisive factor for practical application is low maintenance effort and the ability to change or add to content quickly. That is the only way that chatbots will be able to deliver reliable, up-to-date, intelligent answers during the corona crisis with its multifaceted content and quickly changing rules – and thus help hundreds of thousands of citizens.
Harald Huber ist seit 1991 bei der USU tätig. Er hat beim Aufbau der USU KCenter-Produkte mitgewirkt. Von 2008 bis 2014 war er Produktmanager des Geschäftsbereichs USU KCenter, ab Herbst 2014 bis Ende 2017 bildete er zusammen mit Sven Kolb die Geschäftsleitung USU KCenter. Mit dem Zusammenschluss von vier Geschäftsbereichen der USU, einschließlich KCenter, verantwortet er seit 2018 als Geschäftsführer den daraus entstandenen Bereich unymira, welcher 2021 mit allen USU Solutions unter der Dachmarke USU zusammengeführt wurde. Darüber hinaus ist er langjähriger Autor und Referent für Wissensmanagement-Themen und Trends im Customer Service, sei es Self-Service oder Chatbots.