Get Ready for Chatbots - Part 3
How Do Modern Chatbots Work?
In our previous article we have discussed how insurers can use chatbot to fundamentally reshape and improve their customer interaction. In the following we will give you a short overview of the tech that makes this possible.
As soon as a chatbot receives a message in natural language from a user, it first must identify the user’s intent for sending this message. It then guides the user through the relevant business process by providing him with information and determining required parameters using so-called entities.
Thus, when developing a chatbot it is necessary to first specify the intents and entities which the former must be able to identify. With the help of a few sample user inputs and machine learning, the chatbot is then able to label so far unknown user inputs with the correct intents and entities.
Similarly to a human child learning how to speak, a chatbot initially makes lots of mistakes when processing natural language and gradually improves if it can train and learn from many examples. Due to this analogy, machine learning applications are also often referred to as artificial intelligence.
Let’s illustrate this concept with an example: Tim sends the following message to his insurer’s chatbot to report the damage that has been caused to his windows by a hailstorm:
The intent describes what the user wants to achieve with his request. In our example, the chatbot identifies the intent «File a claim» in Tim’s message:
Entities are used to extract parameters from the message that are required for the business process. For example, the chatbot determines the following entities and associated values in Tim’s message:
The two tables above each contain a column called confidence. It describes how «confident» the chatbot is about recognizing the correct intents and entities in Tim’s message.
The accuracy achieved by a chatbot in identifying intents and entities is improved by training it with different sample user inputs related to the intents and entities. During this so-called supervised learning phase it is checked whether the chatbot has determined the correct results in the examples. If not, the sample user inputs are labeled with the correct intents and entities, and the AI model upon which the chatbot is based on is re-computed.
The more training data a chatbot receives and the more intensely it is trained, the more confident and accurate it becomes. By using a pre-trained NLP model from AI-as-a-service (AIaaS) providers such as Microsoft, Facebook and co., the training effort for a chatbot can be reduced significantly. In most cases, just a few examples already suffice to train a chatbot for a process-specific user input. This allows insurers to build their own chatbots without making the large investments that otherwise would be required to develop AI from scratch.
Now that we understand how modern chatbots work, we will discuss in our next article what aspects should be considered such that insurers can harness their full potential.
You might also be interested in the following articles:
Are you curious about trying out our prototypes? Would you like to talk more about chatbots? You can reach us directly, using the contact details below.