Handling exceptions (incl. ‘query miss’) and hand-overs

The NLP system is not capable of providing automated answers to some users’ utterances. The system has different modes of how to react to requests that are not understood (i.e. ‘don’t knows’). This action is triggered when the predicted score for the answer falls below a certain threshold.

The main fall-back (escalation) options are:

Escalating to a pre-programmed response, optionally with a clarifying scenario
This type of reaction is defined in the same way as a simple response (in the Template window). The suggested topics are questions from the knowledge base entered in the desired order in the Suggested questions window.
From the user's side, it’s visible as a linear dialogue (scenario):

Escalation to a chat with Live Agent

Process of supporting multiple languages

Two approaches as recommended:
  • Providing full NLP with language model, incl. application of full semantic analysis of queries, dictionaries and databases of ready-made responses in a given language This is now available for selected languages, incl English, Polish, Spanish;
  • Machine translation service
Use of a combination of automatic translators and dedicated semantics:
  • Each user input is translated into knowledge base language, e.g. from Spanish into English,
  • then user input is processed internally as usual, but all system outputs are back translated to the user language, e.g. from English into Spanish. Such solution operates only 10-15% less reliable than originally translated knowledge base. There is no time waste for translation services, no time spent on setting up the services - all works on the same knowledge base serviced in one language and serves support in multiple languages (as many as needed).