Heterogeneous Parsing via Collaborative Decoding

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1344–1352, Beijing, August 2010 Heterogeneous Parsing via Collaborative Decoding Muhua Zhu Jingbo Zhu Tong Xiao Natural Language Processing Lab. Northeastern University zhumuhua@gmail.com {zhujingbo, xiaotong}@mail.neu.edu.cn Abstract There often exist multiple corpora for the same natural language processing (NLP) tasks. However, such corpora are gen- erally used independently due to distinc- tions in annotation standards. For the pur- pose of full use of readily available hu- man annotations, it is significant to simul- taneously utilize multiple corpora of dif- ferent annotation standards. In this pa- per, we focus on the challenge of con- stituent syntactic parsing with treebanks of different annotations and propose a col- laborative decoding (or co-decoding) ap- proach to improve parsing accuracy by leveraging bracket structure consensus be- tween multiple parsing decoders trained on individual treebanks. Experimental re- sults show the effectiveness of the pro- posed approach, which outperforms state- of-the-art baselines, especially on long sentences. 1 Introduction Recent years have seen extensive applications of machine learning methods to natural language processing problems. Typically, increase in the scale of training data boosts the performance of machine learning methods, which in turn en- hances the quality of learning-based NLP systems (Banko and Brill, 2001). However, annotating data by human is expensive in time and labor. For this reason, human-annotated corpora are consid- ered as the most valuable resource for NLP. In practice, there often exist more than one cor- pus for the same NLP tasks. For example, for constituent syntactic parsing (Collins, 1999; Char- niak, 2000; Petrov et al., 2006) in Chinese, in ad- dition to the most popular treebank Chinese Tree- bank (CTB) (Xue et al., 2002), there are also other treebanks such as Tsinghua Chinese Tree- bank (TCT) (Zhou, 1996). For the purpose of full use of readily available human annotations for the same tasks, it is significant if such cor- pora can be used jointly. At first sight, a di- rect combination of multiple corpora is a way to this end. However, corpora created for the same NLP tasks are generally built by different orga- nizations. Thus such corpora often follow dif- ferent annotation standards and/or even different linguistic theories. We take CTB and TCT as a case study. Although both CTB and TCT are Chomskian-style treebanks, they have annotation divergences in at least two dimensions: a) CTB and TCT have dramatically different tag sets, in- cluding parts-of-speech and grammar labels, and the tags cannot be mapped one to one; b) CTB and TCT have distinct hierarchical structures. For example, the words “中国 (Chinese) 传统 (tradi- tional) 文化 (culture)” are grouped as a flat noun phrase according to the CTB standard (right side in Fig. 1), but in TCT, the last two words are in- stead grouped together beforehand (left side in Fig. 1). The differences cause such treebanks of different annotations to be generally used in- dependently. This paper is dedicated to solving the problem of how to use jointly multiple dis- parate treebanks …

Content truncated to 500 words