Chi-square assumes that expected counts are greater than 5. However, in larger tables it is allowed to have up to 20% of expected counts less than 5. What is this 20% of? Is it of the cells in the matrix. For example, in a 2x4 table I have 8 cells. Of these 8 cells 1 may have expected counts less than 5 (20% of 8)? Or, am I misunderstanding the assumption?

If I do end up breaking these assumptions I can progress to Fischer's exact test. Is that correctly understood?