diff --git a/.gitignore b/.gitignore
index aedc0ab2e3..aa2206c5f5 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,3 +1,4 @@
 *.pyc
 gensim.egg-info
 *,cover
+.idea
diff --git a/README.git.txt b/README.git.txt
new file mode 100644
index 0000000000..9b764d20b4
--- /dev/null
+++ b/README.git.txt
@@ -0,0 +1,60 @@
+This is my working version of gensim. I keep it synchronized with the upstream
+svn one at assembla.
+I have added some functional tests and utility functions to it. But the main
+reason I'm using the library is to replicate (Gabrilovich & Markovitch, 2006,
+2007b, 2009)  Explicit semantic analisis (ESA).
+
+For other implementations try:
+C#: http://www.srcco.de/v/wikipedia-esa
+java: airhead research library. However the lack of sparse matrix support on
+java linear algebra libraries make java a poor choice.
+
+Currently (as of 27 Aug 2010) , gensim can parse wikipedia from xml wiki dumps quite efficiently.
+However, our ESA code uses a different parsing that we did before (following the
+method section of the paper).
+
+We use here a parsing from March 2008.
+
+Our parsings have three advantages:
+1- THey consider centrality measures, and this is not currently easy to do with
+ the xml dumps directly
+2-
+3- We did an unsupervised name entity recognition parsing (NER) using openNLP.
+THis is parallelized on 8 cores using java code, see ri.larkc.eu:8087/tools.
+We could have used 
+
+NOTE:
+Because example corpora are big, the repository ignores the data folder. Our
+parsing is available online at: (TODO)
+download it and place it under (TODO)
+
+folder structure:
+
+/acme
+    contains my working scripts
+    
+/data/corpora
+    contains corpora.
+
+/parsing
+    tfidf/preprocessing/porter in /parsing adapted from Mathieu Blondel:
+    git clone http://www.mblondel.org/code/tfidf.git
+
+how to replicate the paper
+--------------------------
+code is in /acme/lee-wiki
+
+First you need to create the tfidf space.
+There's a flag. Set createCorpus =  True.
+The corpus creation takes about 1hr, with profuse logging.
+This is faster than parsing the corpus from xml (about 16 hrs) because we do not
+do any xml filtering, stopword removal etc (it's already done on the .cor file).
+
+Once the sparse matrix is on disk, it's faster to read the serialized objects than to
+parse the corpus again.
+
+References:
+------------
+E. Gabrilovich and S. Markovitch (2009) "Wikipedia-based Semantic Interpretation
+for Natural Language Processing", Journal of artificial intelligence research, Volume 34, pages 443-498
+doi:10.1613/jair.2669