The first dialog systems were developed and used as early as the mid-1960s and early 1970s. They were based on rigid rules and followed predetermined procedures. Even today such simple “bots” still play a role when it comes to processing largely standardized processes.
Typical applications include notifying changes of address, entering a SEPA mandate or reporting damage to an insurance company. The questions and answers of the bot system are fixed in such rule-based dialog processes and are usually created manually in the corresponding software. In an expanded form, these simple bots can be connected to subsystems in order to exchange the information collected.
Such dialog systems may be usable in simple, manageable tasks, but they have some serious disadvantages: If, for example, the input of the human dialog partner does not exactly match a stored pattern, the bot cannot do anything with it. Branches and variants in the course of the dialog must be predicted and defined exactly. With every change in the process chain, the dialog branches of the bot must also be adjusted.
Rule-based systems are therefore particularly suitable for processes in which questions and answers are in a direct one-to-one relationship. In more complex scenarios, in which the bot needs additional information from the dialog partner, for example, to be able to answer a question, they fail.
Modern dialog systems are much more flexible. You can communicate with people in natural language and do not need rigid process diagrams. To generate such voice assistants, the following machine learning functions are used, which are mostly implemented using neural networks:
– Recognize and output language: Speech assistants first convert spoken language into text (speech-to-text), analyze it and output the answer in natural-sounding language (text-to-speech).
– Detection of intentions: In order to be able to give an adequate answer, the bot must recognize the content of the (spoken) text and be able to derive the intent of the dialogue partner from it. This analysis is carried out using NLU modules (Natural Language Understanding), which are also based on neural, self-learning networks.
– Find and process information: Based on the identified intentions and the content of a dialog, ML-based systems based on neural networks can independently search for, prepare and make available information in knowledge databases or logistics systems. Answers are not fixed, as is the case with the rule-based variants. Instead, the bot responds flexibly to the question and, if necessary, compiles the answers from various sources. He can also actively ask if important information is still missing to answer the question. If there are several possible solutions, the bot can evaluate them and classify them based on their probability.
– Proactive promotion expansion: On the basis of the dialog history, current bot systems can recognize in which areas there are no answer options or branches, suggest new intents to be modeled and identify missing information.
A requirement analysis should be carried out before the creation of a bot. It defines which specific tasks the digital assistant should take on. To do this, you should collect as much information as possible. For example, existing discussion guides from call centers and support departments or interviews with employees in the relevant departments are helpful.
It is also advisable to work with service providers who have in-depth experience with the industry or department-specific processes to be modeled and can offer pre-trained bots for the respective question.
In addition, it is advisable to coordinate with IT and operations at an early stage and to clarify in which systems the bot has to be integrated and from which sources it obtains information. Legal and regulatory questions must also be addressed. Such requirements can be decisive for the question of whether a bot can be operated in the cloud or not.
Based on the process chains to be implemented and the coordination with the specialist departments, so-called entities are now defined – knowledge units that the bot needs to answer a question and which he may have to actively query. The actual bot development depends on the chosen technology platform. In graphic-based solutions, dialog systems can be put together very quickly even by employees without programming knowledge. They are recommended if users from the specialist departments should continue to look after the bot. Others are based on pure code and can therefore only be operated with appropriate experience.
When the bot is completed in an initial version, the training phase begins. In principle, two forms of learning can be distinguished: The bot can differentiate between good and less good answers based on user reactions and thus increase its hit rate. For this purpose, the feedback should be in an easily interpretable, machine-readable form, for example as a rating star or scale. For free text comments, a human curator is required who translates the answers of the users into actions for bot training.
Alternatively or in addition, the bot can be adapted and improved directly. Based on the previous conversations, experts identify problem areas and process chains in which the bot was unsure, for example, or in which the dialogue was broken off or escalated to a human contact. By analyzing the conversations, specific changes can be made in bot control, such as querying additional information or changing dialog processes.
The goal of process automation is usually that the bot can process a process in a case-by-case manner. While this is unproblematic when ordering a pizza, legal questions arise in the insurance and banking environment, but also in other industries: Who is liable, for example, if a bot makes a wrong decision, for example because it has drawn incorrect conclusions from the training data?
In such cases, it is therefore advisable to use bots to make decisions and to leave the completion of the transaction to a human employee. This has several advantages: on the one hand one avoids the liability discussion, on the other hand the decisions of the bot can be checked directly and if necessary improved by further training. In the best case, this procedure can build up so much trust that the bot can act independently after a transition phase.
Rule-based bots have been known for many years. They work through rigid dialogue requirements and have little to do with AI. Even learning systems based on neural networks make it possible to create voice assistants and chatbots that interact with people in a much more flexible and natural way. These “intelligent” bots recognize questions in natural language, extract a speaker’s intentions and link information from different sources. In this way, the answers of the machine interlocutors are not only getting better, they can also discover new fields of action independently and thus expand their area of application. (mb)