I think you should consider Bag-of-word modelling of messages being generated by the users. You may also go a step further to apply Topic-based modelling (PLSA or LDA) for identifying latent topics of the messages. These modelling step should allow you to group messages based on their contents, hence should allow you to group users based on the content of their messages.

Dear sir

Greetings.

To select the most suitable words from the Bag-of-words, you can use the method of feature selection technique that will solve your problem.

The selection of the words depends on several metrics like statistical measures, variance, frequency, correlation etc. The selection of the metric depends on the researcher.

I hope this helps.

Best regards

Dr.Indrajit Mandal, Ph.D.

thanks, but @Dr. Indrajit Mandal, could you be more specific?  i mean are there any particular technique to do so(to reduce number of words in the bags?), are you aware of similar cases that somehow may the number of features(words in the bags) has been decreased?

Featurs selcted (by bag of words) can be reduced by suitable feature reduction techniques like PCA ,ICA, LDA....

i applied PCA to reduce features obtained by bag of words..given below is the link to my paper..

DOI: 10.1016/j.measurement.2014.04.033

Hassan, - 

take a look into my technical report: https://www.researchgate.net/publication/264834288_Ensemble-based_classification_for_author_profiling_using_various_features

You will find there a comparision of different features

Conference Paper Ensemble-based classification for author profiling using var...

thanks

PLSA or LDA can be used to reduce the dimensions of bag of word document representation.

Extract 1-gram, 2-gram features from the preprocessed text and find out more discriminant feature among all.

Messages need to prioritized based on their ranking. Need to specify one more attribute for that.

I would look for feature like categories, Topics/Themes generated from LDA, hypertext (links are special then words), title. You can use bag of words approach and compute TFIDF of all words in the message as well. 

I think the most distinctive features depends upon your dataset. for example in case of tweets links and hashtags are more important, in case of email messages simple bag of words or topics would be more important. You also need to consider the length of the message. If the length of the message is too large then the message will have more then one topics and using LDA might not help.

If dimensional reduction is an issue, you should at least take a look at Latent Semantic Analysis  - http://lsa.colorado.edu/. I would point out however, that to get the full picture you need to read beyond the academic publications - I found the missing information in their patents.