search

isi-nlp / bolinas

SHERG rule extraction and parsing tools
Other
24 stars 13 forks source link

Generating Grammer #4

Open omidb opened 10 years ago

omidb commented 10 years ago

Hi, Is there any way to generate the SHRG that cover a training set? (given a parallel corpus of graphs, how can I generate the SHRG?) I know that there will be some issues about the definition of "Best Grammar" which I am not that much familiar with.

--Omid

jacobandreas commented 10 years ago

Are you doing string--graph or graph--graph? Do you have alignments? Bolinas has tools for handling both string--graph rule extraction algorithms described in our CoNLL paper. I think the generalization to the graph-graph case should be straightforward.

In the general case you can trivially make rules of the form S -> G_1 : G_2 for every pair (G_1, G_2) of training graphs, but this is probably not what you want.

Somewhat more abstractly, you can always think of synchronously generating graphs as a special case of (non-synchronous) HRG generation where the coindexed nonterminal edges are actually a single hyperedge. You can use this fact to get an upper bound on the treewidth of the grammar necessary to generate your training data. IIRC for a fixed treewidth it's always possible to write down a grammar which generates all graphs of that treewidth.

On Tue, Oct 28, 2014 at 5:17 PM, Omid Bakhshandeh notifications@github.com wrote:

Hi, Is there any way to generate the SHRG that cover a training set? (given a parallel corpus of graphs, how can I generate the SHRG?) I know that there will be some issues about the definition of "Best Grammar" which I am not that much familiar with.

--Omid

— Reply to this email directly or view it on GitHub https://github.com/karlmoritz/bolinas/issues/4.

omidb commented 10 years ago

I'm doing graph--graph, I couldn't find anything to extract SHRGs from data, is there any rule extractor in 'bolinas', if yes, how can I use it?

daniel-bauer commented 10 years ago

Unfortunately the code for rule extraction is currently broken because it is incompatible with the current version of Bolinas. I am planning to fix the old rule extraction code in the near future, as I am starting to work on rule extraction for SHRG and similar formalisms. I can't promise any concrete time frame for this though.

Daniel

Omid Bakhshandeh wrote:

I'm doing graph--graph, I couldn't find anything to extract SHRGs from data, is there any rule extractor in 'bolinas', if yes, how can I use it?

Reply to this email directly or view it on GitHub: https://github.com/karlmoritz/bolinas/issues/4#issuecomment-60860594

omidb commented 10 years ago

Thanks Daniel, please send a note whenever you get it to work.