Chinese Verb Subcategorization Acquisition from Noisy Data on Sentence Level

摘要

Subcategorization is the process that further classifies a syntactic category into its subsets. Aiming to improve the recall of acquisition, we design an automatic approach of enriching the argument knowledge of SCF by means of active learning and employing a multi-class SVM model to classify argument type. We could thus give an accurate SCF as output for each input sentence, even on noisy data, meanwhile avoiding writing rules by hand. Our approach generates hypothesis directly without statistical filtering as the next step after generation. Experiments results indicate that the acquisition performance is significantly improved especially in the aspect of recall, which was increased from 88.83 to 99.75 in open test.