Thank you @Tanja for you advice about the splitting intents. While it is very hard to distinguish from the intents, it is an option. We also applied your idea of doing regular expression prior to each knowledge query, compared with writing re before the interpreter (had to change the re every time when we retrained the interpreter), is much easier.
An example of the control issue, is that when we query entity like a bank, if it is a bank in list A, we act in one way (a dialogue turn) if it is one of the other banks we act in a different way (a different turn). We can represent the dialogue turns in stories, which may confuse the DM model, but when there are too many turns in different situation, we have to write some if statements with some slot altering in the action module, which means we have changed the dialogue flows through the action. And the most hard part in this case is the slot status management. Making some slots empty may suit a situation but gives weird response in the other situation. Just thinking about it makes my head hurt.