Ravikiran Vadlapudi will defend his MSc thesis on Wednesday January 16, 2013.
Vadlapudi, R. Verbose Labels for Semantic Roles.
We introduce a new task that takes the output of semantic role labeling and associates each of the argument slots for a predicate with a verbose description such as buyer or thing bought to semantic role labels such as `Arg0’ and `Arg1’ for predicate like “buy”. Ambiguous verb senses and syntactic alternations make this a challenging task. We adapt the frame information for each verb in the PropBank to create our training data. We propose various baseline methods and more informed models which can identify such verbose labels with 95.2% accuracy if the semantic roles have already been correctly identified. We extend our work to text visualization to illustrate the importance of verbose labeling. As a proof of concept, we built an interactive browser for human history articles from Wikipedia, called lensingwikipedia (http://www.lensingwikipedia.cs.sfu.ca).
The interactive browser for Wikipedia human history articles can be seen live at lensingwikipedia.cs.sfu.ca.