Using a Partially Annotated Corpus to Build a Dependency Parser for Japanese

Using a Partially Annotated Corpus to Build a Dependency Parser for Japanese Manabu Sassano Fujitsu Laboratories, Ltd., 4-1-1, Kamikodanaka, Nakahara-ku, Kawasaki 211-8588, Japan sassano@jp.fujitsu.com Abstract. We explore the use of a partially annotated corpus to build a depen- dency parser for Japanese. We examine two types of partially annotated corpora. It is found that a parser trained with a corpus that does not have any grammatical tags for words can demonstrate an accuracy of 87.38%, which is comparable to the current state-of-the-art accuracy on the Kyoto University Corpus. In contrast, a parser trained with a corpus that has only dependency annotations for each two adjacent bunsetsus (chunks) shows moderate performance. Nonetheless, it is notable that features based on character n-grams are found very useful for a dependency parser for Japanese. 1 Introduction Corpus-based supervised learning is now a standard approach to build a system which shows high performance for a given task in NLP. However, the weakness of such ap- proach is to need an annotated corpus. Corpus annotation is labor intensive and very expensive. To reduce or avoid the cost of annotation, various approaches are proposed, which include unsupervised learning, minimally supervised learning (e.g., [1]), and ac- tive learning (e.g., [2,3]). To discuss clearly the cost of corpus annotation, we here consider a simple model of the cost: annotation cost ∝ � t c(t)n(t) where t is a type of annotation such as POS tagging, chunk tagging, etc., c(t) is a cost per type t annotation, and n(t) is the number of type t annotation. Previous work to tackle the problem of annotation cost has mainly focused on reduc- ing n(t). For example, in active learning, useful examples to be annotated are selected based on some criteria, and then the number of examples to be annotated is considerably reduced. In contrast, we here focus on reducing c(t) instead of n(t). Obviously, if some portion of annotations are not given, the performance of a NLP system will deteriorate. The question here is how much the performance deteriorates. Is there a good trade-off between saving the cost and losing the performance? Minimizing portions of annotations is also very important from the point of view of engineering. Suppose that we want to build an annotated corpus to make a parser for some real-world application. The design and strategy of corpus annotation is crucial in order to get a good parser while saving the cost. Furthermore, we have to keep in mind R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 82–92, 2005. c⃝ Springer-Verlag Berlin Heidelberg 2005 Using a Partially Annotated Corpus 83 the maintenance cost of both the corpus and the parser. For example, we may find some errors in the annotations and the design of linguistic categories. In this situation fewer annotations lead to saving the cost because the corpus is more stable and less prone to errors. The main purpose of this study is to explore the use of a partially annotated corpus to build a dependency parser …

Content truncated to 500 words