Working with syntax trees

[Tae-Gil Noh]

> I also need to devise some general metric that will allow me to estimate
> distance between any two graphs. This distance should account both for
> structural and leaf-node values similarity.
> 
> It would be easier to measure distance between vectors then graphs. So I am
> thinking how to convert directed graph (that results from POS tagging) into
> vector. Any ideas, links here?

R-convolution kernel. R-convolution kernels compare structured data (trees, graphs) and returns a dot-product value. By normalizing this dot-product, you can get a general similarity value. In this case, it seems that you need to compare two parse trees. 

For parse trees: Collins first proposed a parse-tree kernel in the paper Convolution kernels for natural language.

If you need a similarity measure for general graph, instead of tree, there are also graph kernels. I have wrote a few papers using graph kernels on NLP tasks. (http://dl.acm.org/citation.cfm?id=1571941.1572026) 

Tae-Gil Noh 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://projects.haskell.org/pipermail/nlp/attachments/20111029/d7c7dc3f/attachment.htm>