Training your virtual assistant is essential for short-term improvements. Most companies either don't have the right tools or don't have resources planned for them. To accomplish this task, you need at least one employee or even a whole team of intent/content managers, depending on how many conversations the virtual assistant has to handle. The team also needs to be trained and time or resources have to be allocated to deal with the optimization of bots on a daily or weekly basis.

How do virtual assistants work and what is chatbot training?

Every chatbot is built up of intents, which describe different intentions of the users. An Intent could be, for example, that a user wants to know the weather forecast. Intents consist of training phrases and one or more answers. Training phrases, also called utterances, are example questions. Training phrases for the intent "Weather forecast" could therefore be: "What will the weather be like tomorrow?" or "Could you tell me if it's going to rain this weekend?" An answer could be: "The weather will be sunny tomorrow with 20 degrees". Most services used to process and understand natural language work like this. When implementing a chat bot, it is recommended to build an intent with about 15-25 training phrases. This will provide the system with enough data to match user requests with the correct intent. The following applies: The higher the number of question examples, the better the ability of the virtual assistants to answer the users' questions correctly.

Establish processes to train and optimize your virtual assistants 

As soon as the chatbot goes live, training should already have taken place. For this it is crucial to already have a defined process that determines who and when the conversation between bot and user is checked and monitored. This is a continuous process that initially requires more effort and time, while the virtual assistant improves.

The main task is to see if the bot is able to match the right intention with the user's question. If this is the case, the intent manager must confirm the bot's action. Then the user's question is added to the Intent's list of training phrases, if it does not already exist there. If the question was incorrectly assigned, it is necessary to select the correct intent so that the bot enters the question in the list of training phrases. Often chatbots send out error messages indicating that they could not find a suitable intent for the request. In this case it is important to evaluate whether a new intent should be created to add more knowledge to the system. So the next time a similar question appears, the bot will know the answer. This is how chatbot training works.

Use an effective chatbot training tool  

To accomplish this task, the team needs the right tools with specific functionalities. This should be an interface that displays the course of the conversation and the intents matched with the incoming questions. It should be easy to correct or confirm the decision of the bot. It is also very helpful to have filter options to list only conversations within a certain time span, with certain intents or a certain confidence score. The confidence score is a metric from 0 to 1 that indicates how secure the system is with respect to intent detection. For example, you can set the Natural Language Understanding System (NLP) to a score of 0.6. This means that the system will send the response from the intent it has hit when it is at least 60% sure. If the chatbot is only 50% sure that the intent matches the request, the user would receive an error message. In general, the higher the confidence score, the more error messages will occur. The lower the confidence score, the higher the probability that wrong intents will be met. For training purposes, it is then very useful to be able to filter for intents that have been matched with a very low confidence score, as it is more likely to improve the chatbot there. This is especially true for virtual assistants who receive more requests than the team can check. In this case it is recommended to filter the conversation history to work more efficiently.

The Conversational Middleware Platform BOTfriends X offers an interface for training in Intent Management. Learn more about BOTfriends X.

Intent Management Training example

Optimize your virtual assistants in the long term with a strategic and content-related roadmap

If you want to optimize your chatbot, you should consider long-term improvements. Think about which features or additional use cases can be integrated.

For example, the current channel may not be receiving as much traffic or users may not interact with the chatbot as expected. Then you should consider if it makes sense to change channels. For example, from Facebook Messenger to the website. You might also need to rework the overall structure of the conversation by adding new training phrases or removing content.

You also need a roadmap illustrating your "Conversational AI Journey" right from the start. It makes sense to start with a POC, but afterwards scaling is crucial to reach more users. You can roll out to other countries by adding more languages to the bot. Sometimes it is useful to evaluate whether the chatbot has the potential to act as a language assistant or vice versa. Or there are new technologies like NLP services on the market that better fit the needs of your business. In general, many companies start with a simple FAQ bot that answers the most important questions. After evaluating chat data, they plan to add more complex use cases or processes within the chat that require the integration of enterprise systems such as SAP, databases or other software. In summary, these extensions need to be defined in the long-term roadmap in order to pursue a general "conversational strategy". 

If you want to learn more about training your chatbot, you can download our free white paper "Chatbot Operations".