In practice, we use 3, 5 or 7 linguistic variables. It is dependent on the level of distinction. If we use more variables, then we have higher accuracy. Sometimes, we have to think about what accuracy is needed. The best way is query the expert, how many variables are needed to describe the problem.

You do have few choices:

1. Ask experts to let them deliver information on the nature of the problem and possible classification or at least the amount of the variables to partition your universe of discourse (most useful but usually impossible ;-) )

2. Take a look into the training / sample set for the problem and try to find out the local / global maximums / minimums then try i.e. fuzzy-C-means clustering on the training dataset (requires at least a little knowledge on the problem and its nature)

3. Evolve the correct set using GA/ES/EA to let the heuristics find the best suitable set and number of the linguistic variables itself (requires the function to estimate the quality of the classification) - this automated model limits the required knowledge but you do need to be aware that searching full R may be problematics, the limits may be found with ease if you have the training data set (it is advised to extend its min/max values at least a little to let the border curves have the possibility to evolve correctly.