Cascaded Classification for High Quality Head-modifier Pair Selection

This paper presents a cascaded classification approach for selecting head-modifier pairs with high quality from syntactically analyzed sentences. Experimental results show that the proposed approach achieved 76.11% on Fscore of selected head-modifier pairs, which was 8.54% higher than the baseline approach that using sentence length as selection criteria. In addition, compared with using the headmodifier pairs extracted by the baseline approach, the parsing accuracy of a lexicalized dependency parser can be further improved by using the head-modifier pairs selected by the proposed approach.