nn in information retrieval

Hello fuzzy / neural network implementers


i am building an extension for lucene, the search engine project within 
apache , that integrates the user´s
query history into the retrieval process.

one idea is to implement a neural network, that adapts to the users
information need by reweighting search terms.

furthermore there are 2 different levels of text representation, one by 
a large collection of metadata and pure fulltext search, and a second 
based on the same metadata AND the entire document.
the latter is permanently changed in order to obtain a representation 
closer to what a single user considers relevant.

first, the application performs a full-text search on a metadata corpus 
using lucene.

the user then selects the apparently most interesting documents (by its
dc metadata). this second user-input should be used to adapt the neural
network weights.


until now, i am not entirely decided which neural network api i should
use. joone, at a first glance, seems to be quite user friendly.

question 1 :

does an easy to use "term-document-matrix builder" exist, that 
constructs t-dimensional matrix containing all the terms in the corpus 
mapped to the documents where they occured?
it would be helpful if it were easily possible to compute the 
tf-idf-measure instead of the absolute term-occurence values (integers).

i thought of a simple 2-layered nn, in wich one of the layers represents 
documents and the second one all the terms in the corpus (cf. google: 
cosimir)

so, once the users decides to see the full text of several docs 
refenrenced in metadata retrieved by a full text search according to teh 
users initial query terms, it must be possible to increase the 
activation of those neurons that represent the documents selected for 
full retrieval.

quetion 2:

do you know whether there is a more appropriate (open source) software 
to build such a nn?

i am a real newbie to ai and appreciate your opinion.


thanks for any hint!


thomas krämer