Weakly supervised training for parsing Mandarin broadcast transcripts

We present a systematic investigation of applying weakly supervised co-training approaches to improve parsing performance for parsing Mandarin broadcast news (BN) and broadcast conversation (BC) transcripts, by iteratively retraining two competitive Chinese parsers from a small set of treebanked data and a large set of unlabeled data. We compare co-training to self-training, and our results show that performance using cotraining is significantly better than with self-training and both co-training and self-training with a small seed labeled corpus can improve parsing accuracy significantly over training on the mismatching newswire treebank. We also investigate a variety of example selection approaches for co-training and find that our proposed example selection approach based on maximizing training utility produces the best parsing accuracy. We also investigate Chinese parsing related issues including characterbased parsing and word segmentation for parsing.