try spectral clustering, LDE, etc.

look at the dimensionality reduction toolbox for matlab (3rd party)

If you want key word sets for tagging then two possibilities suggest themselves; both use what is called the "Distribution Hypothesis", that linguistic items with similar distributions have similar meanings.

The difference between the two is that Word2Vec positions the words, but not the sets of words. The second approach positions the sets, but not the words, however it needs nifty logic, and stop words  to avoid an exponential explosion. Stopwords?  Why count 'and' or 'the' if Google doesn't?

Fortunately both approaches are available in Python, so you can compare directly for yourself using the attached Jupyter notebook. My take is that counting provides the gold standard, but both support the hypothesis that Firth articulated just sixty(!) years ago, "You shall know a word by the company it keeps".

https://en.wikipedia.org/wiki/John_Rupert_Firth

Hi Deepanshu, 

You could use PCA to extract the features. After that, cluster those features using a clusterer method (e.g., K-means). At this stage, you may need to test several distance functions in order to cluster your data. Bear in mind that using an incremental learning with a small dataset, might not be a good idea.  

HTH.

Samer