U. Baldewein and K. Erk and S. Pado and D. Prescher: Semantic Role Labelling for Chunk Sequences. Proceedings of the CoNLL'04 shared task, Boston.
Note: The results in this paper were obtained from a system which turned out to be somewhat buggy; the judgements about the (limited) usefulness of clustering for generalisation should therefore be reconsidered.
We describe a statistical approach to semantic role labelling that employs only shallow information. We use a Maximum Entropy learner, augmented by EM-based clustering to model the fit between a verb and its argument candidate. The instances to be classified are sequences of chunks that occur frequently as arguments in the training corpus. Our best model obtains an F score of 51.70 on the test set.
@InProceedings{baldewein04:_seman_role_label_chunk_sequen,
author = {U. Baldewein and K. Erk and S. Pado and D. Prescher},
title = {Semantic Role Labelling for Chunk Sequences},
booktitle = {Proceedings of the CoNLL'04 shared task},
year = 2004,
address = {Boston, MA}
}