Why a virtual assistant detects the intent, even if there are no such words and meanings in the intent examples
Virtual assistant trainers often ask why the intent detector detects the intent for phrases having a different meaning.
The intent detector does not understand the meaning of the text but takes it purely robotically by analyzing letter combinations and stacks from what it has learned against the text entered by the user. It then calculates the confidence: how high is the confidence that the text entered by the user corresponds to one or the other intent in the knowledge of the assistant.
For example, let's imagine that the virtual assistant knows the intent A, which includes a variety of short fairy tale names “Snow White", "Sleeping beauty", "Puss in Boots", and other similar names. When the user asks something about the food "Vegetable stew", "Pavlova cake" and "Herring in oatmeal", a virtual assistant is more likely to detect the intent A.
A virtual assistant usually has its purpose and knowledge of one field. When users ask about unrelated topics about which a virtual assistant is not trained, of course, an intent from an existing knowledge base may be detected.
If this happens, assess how much work to invest in such cases. If it appears that users ask a lot about the topic, and it is important that the virtual assistant answers correctly, you can:
- create a separate intent that provides a necessary response (using the virtual assistant's tone of voice);
- create a general answer: “I know about... Choose!" and add here to the intent all those cases that are disturbing.
The intent detector works better when intent examples consist of full sentences and questions, not a single word.