I believe it is because of the nature of the set that we are attempting to classify or describe. The letters are symbols could be assembled into an object — language can be thought of in a similar sort of way. So think of legos, we attempt to classify the lego pieces as ‘a’ or ‘b’ piece. And the group of legos is as such that certain pieces go together and certain ones do not. It’s the structure of the pieces that cause this behaviour. So we find that we can attach a and b pieces together b and a pieces together and ‘a’ and ‘a’ and ‘b’ and ‘b’. But when we try to add another piece we find that there is no place when xx exists that we can attach a ‘b’ piece to make the object xxb which mean anything. Such is the structure of the said pieces and therefore it’s an impossible combination. Now we have described a set of said pieces. If we are presented with another bag of pieces and we attempt to create xxb and it works, then the description that we had of another set does not apply because in the other set such combinations are impossible. This just adds rigour to your description that you would otherwise not have.

If it’s a descriptive system or I prefer the word model then it should have a structure such that some combination of symbols is impossible (so that the classification system limits certain things and has a certain limited structure) so that if we look at some set of data later that is said to follow our model and it does not have the same linguistic structure then we can say that this data does not fit this model and therefore assert that the model is not valid for the given data.

So say we have a bag of words that has words made of the letters a and b, the length of the word is either 2, 3, 4, 5 letters. We pick out all the 2 letter words to study and find that all combinations are possible, so ab ba aa bb. Then we do the same thing for the 3 letter words and find that the words never end in b, so aaa, aba baa bba is all that we find. We have observed an impossibility which says something about the structure of the data that we are studying. We could continue and stud the 4 and 5 letter words. But the fact that we have learned that abb is impossible has said that one combinations of letter is impossible in our data. If we include this fact in our descriptive model then we can latter assert for another set of data that has abb is not described by our model.

Referenced from:
http://books.google.ca/books?id=1Zw9AAAAIAAJ&lpg=PA24&dq=impossible%20combination%20of%20symbols&pg=PA25#v=onepage&q=impossible%20combination%20of%20symbols&f=false