Analyzing the Effect of Global Learning and Beam-Search on Transition-Based Dependency Parsing

Beam-search and global models have been applied to transition-based dependency parsing, leading to state-of-the-art accuracies that are comparable to the best graph-based parsers. In this paper, we analyze the effects of global learning and beam-search on the overall accuracy and error distribution of a transition-based dependency parser. First, we show that global learning and beam-search must be jointly applied to give improvements over greedy, locally trained parsing. We then show that in addition to the reduction of error propagation, an important advantage of the combination of global learning and beam-search is that it accommodates more powerful parsing models without overfitting. Finally, we characterize the errors of a global, beam-search, transition-based parser, relating it to the classic contrast between “local, greedy, transition-based parsing” and “global, exhaustive, graph-based parsing”. TITLE AND ABSTRACT IN CHINESE 分析全局模型和柱搜索对基于转移依存分析器的影响 柱搜索和全局模型被应用于基于转移的依存分析,可以取得与最好的基于图的依存分析器 同一水平的精度。我们分析全局学习和柱搜索对基于转移的依存分析器的精度与错误分布 的影响。首先,全局学习和柱搜索需要同时使用才能达到显著优于局部学习和贪婪搜索的 效果。此外,全局学习和柱搜索的联合使用不仅可以减少错误蔓延,还可以支持更为复杂 的模型训练而不过拟合。最后,我们对应用了全局学习和柱搜索的基于转移的依存分析器 进行错误分析,且将此分析与对MaltParser与MSTParser的错误对比相比较。