Experiments in Cross-Language Morphological Annotation Transfer

Annotated corpora are valuable resources for NLP which are often costly to create. We introduce a method for transferring annotation from a morphologically annotated corpus of a source language to a target language. Our approach assumes only that an unannotated text corpus exists for the target language and a simple textbook which describes the basic morphological properties of that language is available. Our paper describes experiments with Polish, Czech, and Russian. However, the method is not tied in any way to these languages. In all the experiments we use the TnT tagger ([3]), a second-order Markov model. Our approach assumes that the information acquired about one language can be used for processing a related language. We have found out that even breathtakingly naive things (such as approximating the Russian transitions by Czech and/or Polish and approximating the Russian emissions by (manually/automatically derived) Czech cognates) can lead to a significant improvement of the tagger’s performance.

标注语料库是自然语言处理的宝贵资源，创建它们往往成本高昂。我们介绍一种将标注从源语言的形态标注语料库转移到目标语言的方法。我们的方法仅假定目标语言存在未标注的文本语料库，并且有一本描述该语言基本形态属性的简单教科书。我们的论文描述了针对波兰语、捷克语和俄语的实验。然而，该方法与这些语言没有任何特定的关联。在所有实验中，我们使用TnT标注器（[3]），这是一个二阶马尔可夫模型。我们的方法假定关于一种语言获取的信息可用于处理相关语言。我们发现，即使是极其简单的方法（比如用捷克语和/或波兰语来近似俄语的转移概率，以及用（手动/自动推导的）捷克同源词来近似俄语的发射概率）也能显著提高标注器的性能。