The Treehouse presents:
Rebecca Dridan, U. Melbourne
October 15, 2010 12:30-1:20 pm
In many cases, NLP is a consumer of the results of grammar engineering, but NLP methods can also be used to contribute to the grammar engineering process. At the University of Melbourne, we currently have a number of projects aimed at assisting grammar engineers. In this talk, I am going to present details of two of them:
MUPS: The aim of this project is to provide a means to build a statistical parse selection model without the manual treebanking stage. Our first attempt uses supertag sequences to direct the Maximum Entropy estimation towards preferred analyses. Experiments using this model achieved parsing accuracy significantly better than any unsupervised baseline for both Japanese and English data, providing a potential method to bootstrap models for new grammars without treebanks, and also possibly a cheap method for domain adaptation for more mature grammars.
gDelta: This tool uses data mining methods to assess the impact of a grammar change at a more fine-grained level than the counts of items parsed and readings that the standard treebank profiling tool reports. It provides different visualisations of a change, focussing on changes in which rules were applied. These visualisations can be used to confirm that a change was successful, but also to identify unexpected side-effects caused by interactions in a complex grammar.
We welcome any feedback on the usefulness of these tools and on what other features or information could be helpful.
Topic revision: r1 - 2010-10-07 - 05:16:21 - ebender