Add support for normalizing RaKUn's output based on tokenized text
RaKUn's keywords are based on lemmatized input and thus some resulting keyphrases might sound unnatural ("majandusteaduskond dekaan"). To overcome this problem, align lemmas with tokenized text and find the original form of the first word in the phrase (the second word should be lemmatized). Although there are some exceptions on which this solution doesn't work (and makes the output even more unnatural), it will hopefully improve the results in most cases.