Related approaches for analysis of social networks:

http://www.pbs.org/newshour/rundown/mapping-twitter-conversations-social-structures/

http://arxiv.org/abs/1211.6086

http://www.cmu.edu/joss/content/articles/volume8/Welser/

https://www.technologyreview.com/s/420648/how-to-tell-who-is-influencing-whom-in-a-group-discussion/

http://www.isi.edu/natural-language/people/hovy/papers/06HLT-NAACL-thread-focus.pdf

I think an expert knows synonyms like platelets and thrombocytes

and might use them in vicinity, but for detecting these

you need a specific lexicon.

Regards,

Joachim

Agree,these might be useful.

https://www.technologyreview.com/s/420648/how-to-tell-who-is-influencing-whom-in-a-group-discussion/

http://arxiv.org/abs/1211.6086

Dear Marouane Soula,

If you have large data set to train a classifier, I suggest you use a supervised text categorizer/classifier.

If you have specific ontology for each domain then you can compare it to the ontology of the written text. This method needs ontology learning. You can choose the specific ontology by using TF-IDF.

best regards,

Harry

Opinion spam detection may assist you in this regard. number of articles are available , just Google them

Yes. Opinion spam detection will really assist you in this area. You can also base your work on Point-wise Mutual Information (PMI) or use other sophisticated semantic similarity measure (https://en.wikipedia.org/wiki/Semantic_similarity) to address it. The only issue is that your work will be domain dependent, which means that you need domain dependent lexicons to do the mining and detection of specialist domain specific words.

All the best.

I think your problem can be classified as binary classification. The first option is to use ontology to identify if these words are domain specific or generic words. The second option is to use topic modeling techniques such as TF-IDF  but this will face the challenge of the short text, so you may need to use the Bi-term topic modeling from the following url: 

http://delivery.acm.org/10.1145/2490000/2488514/p1445-yan.pdf?ip=129.127.137.193&id=2488514&acc=ACTIVE%20SERVICE&key=65D80644F295BC0D%2E001A23AA3BABC648%2E4D4702B0C3E38B35%2E4D4702B0C3E38B35&CFID=753167755&CFTOKEN=99842272&__acm__=1455524863_98545fa51dbd30a8ac021d86aec3392b 

you can also check this question:  https://www.researchgate.net/post/What_is_a_good_way_to_perform_topic_modeling_on_short_text

the third option if you have large amount of data is to train a SVM to recognise the difference between the experts and non experts. I expect the SVM to recognise the named entities that are associated with professionals