Vaelen

Update on Selectional Preferences

Posted by Andrew Mon, 02 Mar 2009 08:57:00 GMT

It’s looking more and more like the addition of selectional preferences is actually causing a decrease in classifier performance in my experiments. I’ve tweaked the set of “bins” for a given cosine to a small set: zero, small, large, exact. The “small” bin denotes that the cosine was between 0 and .5, and the “large” bin denotes that the cosine was between .5 and 1. In most cases, the performance of the new dataset is either the same or slightly worse than the performance of the benchmark dataset.

Here are some examples These are being run against my training and development datasets and the Judgment_communication frame is the largest dataset I have to work with. The J48 tree learner seems to get more out of the addition of selectional preferences than the SMO learner.

Initial Results of Selectional Preferences (Abundance Frame)

Posted by Andrew Thu, 26 Feb 2009 07:42:00 GMT

I’m testing whether or not adding selectional preferences improves the ability of a classifier to properly assign semantic role labels. To generalize word meaning I create a highly dimensional vector space from a dependency parse of the text. Selectional preferences are represented by comparing the cosine of the angle between the given words vector and the vectors of all the labeled instances from the training data.

This result was generated from a single frame (Abundance). The next step is to run this across the entire FrameNet corpus and compile the results.