Transfer learning for event noun SRL.

In our paper (Pado, Pennacchiotti and Sporleder COLING 2008), we have argued that it is possible to use verbal SRL training data to label the semantic roles of corresponding event nominalizations.

The file that is downloadable from here specifies the dataset that we have used for our evaluation. It is a five-column, tab-separated text file:

  • Column 1: nominalisation (e.g., announcement)
  • Column 2: verb (e.g., announce)
  • Column 3: frame (e.g. Statement)
  • Column 4: FrameNet Lexical Unit filename for noun
  • Column 5: FrameNet Lexical Unit filename for verb

Each line in this file thus specifies one frame-disambiguated verb-noun pair for which we used the verbal data to predict the nominal data and for which our preprocessing with the Collins parser yielded at least one verbal and one nominal instance. This dataset is composed of 238 unique nouns, 235 unique verbs, and 117 unique frames. We used FrameNet release 1.3.


Feedback is always welcome at