self-training with products of latent variable grammars.ppt

A Closer Look at Regular Results * A Closer Look at Regular Results * A Closer Look at Regular Results * A Closer Look at Self-Training Results * A Closer Look at Self-Training Results * A Closer Look at Self-Training Results * Analysis of Rule Variance We measure the average empirical variance of the log posterior probabilities of the rules among the learned grammars over a held-out set S to get at the diversity among the grammars: * Analysis of Rule Variance * English Test Set Results (WSJ 23) Single Parser Reranker Product Parser Combination [Charniak ’00] Petrov et al. ’06] [Carreras et al. ’08] [Huang Harper ’08] This Work [Petrov ’10] This Work [Charniak Johnson ’05] [Huang ’08] [McClosky et al. ’06] [Sagae Lavie ’06] [Fossum Knight ’09] [Zhang et al. ’09] * Broadcast News * Conclusions Very high parse accuracies can be achieved by combining self-training and product models on newswire and broadcast news parsing tasks. Two important factors: Accuracy of the model used to parse the unlabeled data Diversity of the individual grammars * 1. Overview of latent variable PCFG grammars: two problems with EM training a. Variation: simple solution: use dev set better solution: product model b. Overfitting: simple solution: smoothing better solution: self-training 2. Problems with product models and self-training a. Product model: overfitting still exits in the individual grammars, but it turns out not to be a problem in the product model grammar size becomes large and parsing becomes slow b. Self-training: variation still remains 3. Joint solution using both self-training and product models a. training on the output of a product model results in individual grammars that are as good as the product: smaller memory footprint and faster decoding b. product on self-trained grammars c. diversity matters. * Re