Training Dependency Parsers from Partially Annotated Corpora

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 776–784, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Training Dependency Parsers from Partially Annotated Corpora Daniel Flannery1, Yusuke Miyao2, Graham Neubig1, Shinsuke Mori1 1Graduate School of Informatics, Kyoto University Yoshida Honmachi, Sakyo-ku, Kyoto, Japan 2National Institute of Informatics 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo, Japan flannery@ar.media.kyoto-u.ac.jp, yusuke@nii.ac.jp, neubig@ar.media.kyoto-u.ac.jp, forest@i.kyoto-u.ac.jp Abstract We introduce a maximum spanning tree (MST) dependency parser that can be trained from partially annotated corpora, allowing for effective use of available lin- guistic resources and reduction of the costs of preparing new training data. This is especially important for domain adapta- tion in a real-world situation. We use a pointwise approach where each edge in the dependency tree for a sentence is es- timated independently. Experiments on Japanese dependency parsing show that this approach allows for rapid training and achieves accuracy comparable to state-of- the-art dependency parsers trained on fully annotated data. 1 Introduction Parsing is one of the fundamental building blocks of natural language processing, with applications ranging from machine translation (Yamada and Knight, 2001) to information extraction (Miyao et al., 2009). However, while statistical parsers achieve higher and higher accuracies on in-domain text, the creation of data to train these parsers is labor-intensive, which becomes a bottleneck for smaller languages. In addition, it is also a well known fact that accuracy plummets when tested on sentences of a different domain than the train- ing corpus (Gildea, 2001; Petrov et al., 2010), and that in-domain data can be annotated to make up for this weakness. In this paper, we propose a maximum span- ning tree (MST) parser that helps ameliorate these problems by allowing for the efficient develop- ment of training data. This is done through a com- bination of an efficient corpus annotation strategy, and a novel parsing method. For corpus construc- tion, we use partial annotation, which allows an annotator to skip annotation of unnecessary edges, focusing their efforts only on the ones that will provide the maximal gains in accuracy. While partial annotation has been shown to be an effective annotation strategy for a number of tasks (Tsuboi et al., 2008; Sassano and Kurohashi, 2010; Neubig and Mori, 2010), traditional MST parsers such as that of McDonald et al. (2005) can- not be learned from partially annotated data. The reason for this is that they use structural predic- tion methods that must be learned from fully an- notated sentences. However, a number of recent works (Liang et al., 2008; Neubig et al., 2011) have found that it is possible to ignore structure and still achieve competitive accuracy on tasks such as part-of-speech (POS) tagging. Similarly, recent work on dependency parsing (Spreyer and Kuhn, 2009; Spreyer et al., 2010) has shown that training constraints can be relaxed to allow MST parsers to be trained from partially annotated sentences, with only a small reduction in parsing accuracy. In this approach the scor- ing function used to evaluate potential dependency trees is modified so …

Content truncated to 500 words