I code for a mud and having started to study neural networks and information retrieval, I got to thinking that as an experiment, a neural network could be used to create an improved help file system through an mob. So I was wondering if I could create a backprop NN and train it to associate certain help files to certain queries? I am at a loss though as to how I can encode the information into a network. I know some researchers have tried to create a binary matrix of keywords and document ids, with 1 or 0 indicating the keyword is present within the document eg ____________doc1 doc2 doc3 fireball_____1 0 0 combat_______1 1 0 kick_________0 0 0 but I would imagine that such an approach would require an large and constantly updated NN. Any ideas/comments? [edited by - theObserver on March 12, 2004 4:49:40 PM] [edited by - theObserver on March 12, 2004 4:50:39 PM]

While you might be able to get this to work, I think it will be difficult. Most artificial neural networks are best at learning to interpolate among example data, and I''m not sure how well documents fit this paradigm. Research text-specific algorithms, like the j-measure.


-Predictor
http://will.dwinnell.com

What kind of queries are you expecting, freeform natural language ones? And how would you train it - ask the players if it answered their question or not? I''m not sure this would be the best way of implementing this.

[ MSVC Fixes | STL Docs | SDL | Game AI | Sockets | C++ Faq Lite | Boost
Asking Questions | Organising code files | My stuff | Tiny XML | STLPort]

In the information retrieval field, systems use simple boolean matching ie user enters a natural language search query and the system strips out the common words like "is" or "the". The remaining words are then matched against the documents and a list is returned of all documents found containing those words, ranked in order of relevance to the users query depending on the proximity of the keywords to each other.

However researchers are experimenting with AI techniques such as neural networks, generic algorthims and NLP to create an more intelligent IR system cabable of ''understanding'' what information the user wants if a vague query is inputted.

I am studying this in college and as a general interest side-project decided to have a shot at implementing a NN based solution in my mud (although I 100% agree it is not the best or even a practical solution).

One theory is that similar queries will have similar relevant documents. So a NN''s ability to generalize could be used to associate queries to documents. To train the NN (again in theory) a query is inputted into the input layer with the output layer set to the correct relevant documents and a back prop algorthim used. The player can further train the NN by simply saying ''yes'' or ''no'' to each help file returned and the NN weights are adjusted accordingly.

I am probarly grasping at straws here but anyone got any idea of how to encode the input/outputs to implemenate such a NN?