Seminar on Computational Learning and Adaptation


Learning to Identify Semantic Roles

Cindi Thompson
PriceWaterhouseCoopers

Identifying and classifying the key semantic components of a sentence or phrase  is an important part of determining sentence meaning.  We present a model of natural language generation from semantics using the FrameNet semantic role and frame ontology.  We present a learning approach for identifying semantically relevant components of a syntactic parse, using its class labels to help train a generative model for role labeling.  The model is trained in a novel manner to identify both roles and irrelevant components.  Our model makes fewer assumptions than prior models applied to the task, and it produces results competitive with previous work.

Portions of this research were done jointly, with Chris Manning and others, when the speaker was a Visiting Professor at Stanford.


Date: Wednesday, February 9, 2005

Time: 4:15-5:30PM

Place: Gates 104


Co-sponsored by the Natural Language and Speech Processing Colloquium.

Return to the seminar schedule