next up previous
Next: Grammar-based NLP Up: Data Oriented Parsing Previous: Methods for word graphs.

Methods for test sentences.

For sentences, we used the same parameter settings, but added an experiment with d=5.

In this document, some results on sentences are given with the extra indication group. We will now briefly explain what this means.

Non-terminals in the semantic DOP-model consist of a syntactic-category / semantic-type pair. Such a non-terminal imposes a rather rigid constraint on substitution. For the parsing of word graphs, this constraint seems to be beneficiary. For the parsing of sentences, on the other hand, these constraints could be too rigid. A greater degree of freedom results in over-generation, which in turn may lead to better statistics. An algorithm was designed to group semantic types that have a comparable distribution. This results in fewer non-terminals in the tree-bank, and has been shown to lead to a higher semantic accuracy for sentences. The results marked with group indicate that this grouping algorithm has been employed.