Iterative Annotation Transformation with Predict-Self Reestimation for Chinese Word Segmentation

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 412–420, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Iterative Annotation Transformation with Predict-Self Reestimation for Chinese Word Segmentation Wenbin Jiang and Fandong Meng and Qun Liu and Yajuan L¨u Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences {jiangwenbin, mengfandong, liuqun, lvyajuan}@ict.ac.cn Abstract In this paper we first describe the technol- ogy of automatic annotation transformation, which is based on the annotation adaptation algorithm (Jiang et al., 2009). It can auto- matically transform a human-annotated cor- pus from one annotation guideline to another. We then propose two optimization strategies, iterative training and predict-self reestimation, to further improve the accuracy of annota- tion guideline transformation. Experiments on Chinese word segmentation show that, the it- erative training strategy together with predict- self reestimation brings significant improve- ment over the simple annotation transforma- tion baseline, and leads to classifiers with sig- nificantly higher accuracy and several times faster processing than annotation adaptation does. On the Penn Chinese Treebank 5.0, it achieves an F-measure of 98.43%, signif- icantly outperforms previous works although using a single classifier with only local fea- tures. 1 Introduction Annotation guideline adaptation depicts a general pipeline to integrate the knowledge of corpora with different underling annotation guidelines (Jiang et al., 2009). In annotation adaptation two classifiers are cascaded together, where the classification re- sults of the lower classifier are used as guiding fea- tures of the upper classifier, in order to achieve more accurate classification. This method can automat- ically adapt the divergence between different an- notation guidelines and bring improvement to Chi- nese word segmentation. However, the need of cas- caded classification decisions makes it less practical for tasks of high computational complexity such as parsing, and less efficient to incorporate more than two annotated corpora. In this paper, we first describe the algorithm of automatic annotation transformation. It is based on the annotation adaptation algorithm, and it focuses on the automatic transformation (rather than adapta- tion) of a human-annotated corpus from one annota- tion guideline to another. First, a classifier is trained on the corpus with an annotation guideline not de- sired, it is used to classify the corpus with the an- notation guideline we want, so as to obtain a corpus with parallel annotation guidelines. Then a second classifier is trained on the parallelly annotated cor- pus to learn the statistical regularity of annotation transformation, and it is used to process the previous corpus to transform its annotation guideline to that of the target corpus. Instead of the online knowl- edge integration methodology of annotation adapta- tion, annotation transformation can lead to improved classification accuracy in an offline manner by using the transformed corpora as additional training data for the classifier. This method leads to an enhanced classifier with much faster processing than the cas- caded classifiers in annotation adaptation. We then propose two optimization strategies, iter- ative training …

Content truncated to 500 words