I am trying to build an intent classifier using RASA. But in addition to the natural language text input I also need to pass in additional information like previously identified intents and other domain specific numeric data which also influences the actual intent. The intent may not be completely clear by just analysing the natural language sentence. Now i want the system to learn the relationships between them(the previous intent and numerical data with the current natural language sentence) instead of having a rule based one.
Is there a way to pass diverse input data to RASA ? I am under the impression that the pipelines in RASA are sequentially chained and the output of one component is fed as the input into the next. Is there a concept of parallel input featurizer pipeline ? I want the natural language sentence to follow the ConveRTTokenizer -> ConveRTFeaturizer path. But i dont want the 'previous intent data" and the numerical data to follow that path. I want all these various vector representations to be finally concatenated as the DIET classifier input.
Thanks,