Omega Ontology: Verb frames component

An important addition to Omega for this release is a set of verb frame annotations. Several verb frame corpora have been harmonized as part of the ontology annotation effort at USC/ISI. These include verb frames from FrameNet, Propbank, and the LCS theta grid/semantic roles corpus. Also included in this section are the annotations to WordNet's syntactic verb frames.

We include here the background and rationale for this verb frame annotation activity.

To give semantic analysis based on the predicate in a sentence, we assign frame information for each word sense of the predicate. There are available data collections which define the semantic frame structures having the predicates: FrameNet, PropBank, and LCS database. Each has a different view and a different coverage, so we integrate all information into the Omega ontology. We currently assign frame information only to verb senses and align frame roles among frames.

FrameNet defines semantic frames involving various participants, properties, and other conceptual roles, and for each frame, corresponding words are associated. FrameNet II(Jan.2004) defines 487 distinct frames and 6,743 predicate lexicons (2,300 verbs) are associated with the frames.

PropBank defines predicate-argument structure on a per-predicate basis, and the core elements of each predicate are simply numbered. PropBank (Feb.2004) covers 3,323 predicate verbs and 4,659 framesets.

LCS is a hand-tagged data on a per-class basis based on the Levin's English verb classes and alternations and it contains 4,452 verbs in 492 classes.

After we assign each sense PropBank and LCS frame structures manually, we assign FrameNet frame and align frame roles among FrameNet, Propbank, and the LCS database. Using the concept relations in Omega and the frame descriptions in each collection, we align frame roles between FrameNet and PropBank, and between FrameNet and LCS. To make more accurate data, we check the alignment manually and correct them.