Friday, March 30, 2007

SemEval 2007

Picture: Christmas 2005, Rachel's Aunt Amy's house in northwestern Ohio.
From left: Erin Drum(Rachel's sister), me, Rachel Drum(my fiance), Kevin Drum(Rachel's brother), Pam Drum(Rachel's Mom), Bruce Drum(Rachel's Dad)

I thought you might be interested to hear about a project I've been working on :-)

This semester I'm taking a seminar in which we are competing in a sort of academic task called SemEval 2007. SemEval offers the oppourtunity for participants(mostly teams at universities, like ours) to attempt a variety of tasks related to automated semantic evaluation. The professor who is teaching the course, Roxana Girju, helped to organize one of the tasks, and so of course we are working on that one. It is task 4, "Classification of Semantic Relations between Nominals".

So, what does that entail? Basically, we're given a bunch of sentences with two nouns marked on each sentence, and most of the time the exact meanings of the marked nouns. We must write a computer program that automagically determines whether or not the two marked nouns have a particular semantic relation. There are seven semantic relations that we have to sort out, so it's not too bad. Here's an example of what we get:

001 "The period of [e1]tumor shrinkage[/e1] after [e2]radiation therapy[/e2] is often long and varied (mean months)."
WordNet(e1) = "shrinkage%1:11:00::", WordNet(e2) = "radiation_therapy%1:04:00::", Cause-Effect(e2,e1) = "true", Query = "* after radiation therapy"

The first line is the sentence, which has "tumor shrinkage" and "radiation therapy" marked. The second line is a reference to the exact Wordnet definition, or "sense", of each marked noun. It also displays the relevant semantic relation, Cause-Effect, which in this case is labeled as a true relation between the two nouns. That is, radiation therapy causes tumor shrinkage, which is straightforward enough for anyone to understand, and this is what is meant by Cause-Effect. The Query indicates a Google search that led the task moderators to this sentence, which comes from the web.

So what's the point of doing this task? A primary application is as an aid to question answering computer programs. For example, suppose you're a doctor and you want to know what shrinks tumors. Suppose there's this amazing program where the doctor can just type in a question and out comes the answer: "radiation therapy". In order to get this answer, the program can search the web, find a sentence like the one given above, and somehow figure out that "radiation therapy" is in a Cause-Effect relationship to "tumor shrinkage". To figure out this "somehow" is the goal of this task.

The deadline for submitting our results is Sunday, so I'll let you know how we do!

No comments: