Dependency Parsing with Short Dependency Relations in Unlabeled Data

Dependency Parsing with Short Dependency Relations in Unlabeled Data Wenliang Chen, Daisuke Kawahara, Kiyotaka Uchimoto, Yujie Zhang, Hitoshi Isahara Computational Linguistics Group National Institute of Information and Communications Technology 3-5 Hikari-dai, Seika-cho, Soraku-gun, Kyoto, Japan, 619-0289 {chenwl, dk, uchimoto, yujie, isahara}@nict.go.jp Abstract This paper presents an effective dependency parsing approach of incorporating short de- pendency information from unlabeled data. The unlabeled data is automatically parsed by a deterministic dependency parser, which can provide relatively high performance for short dependencies between words. We then train another parser which uses the informa- tion on short dependency relations extracted from the output of the first parser. Our pro- posed approach achieves an unlabeled at- tachment score of 86.52, an absolute 1.24% improvement over the baseline system on the data set of Chinese Treebank. 1 Introduction In dependency parsing, we attempt to build the dependency links between words from a sen- tence. Given sufficient labeled data, there are sev- eral supervised learning methods for training high- performance dependency parsers(Nivre et al., 2007). However, current statistical dependency parsers pro- vide worse results if the dependency length be- comes longer (McDonald and Nivre, 2007). Here the length of a dependency from word wi and word wj is simply equal to |i − j|. Figure 1 shows the F1 score1 provided by a deterministic parser rela- tive to dependency length on our testing data. From 1precision represents the percentage of predicted arcs of length d that are correct and recall measures the percentage of gold standard arcs of length d that are correctly predicted. F1 = 2 × precision × recall/(precision + recall) the figure, we find that F1 score decreases when de- pendency length increases as (McDonald and Nivre, 2007) found. We also notice that the parser pro- vides good results for short dependencies (94.57% for dependency length = 1 and 89.40% for depen- dency length = 2). In this paper, short dependency refers to the dependencies whose length is 1 or 2. 30 40 50 60 70 80 90 100 0 5 10 15 20 25 30 F1 Dependency Length baseline Figure 1: F-score relative to dependency length Labeled data is expensive, while unlabeled data can be obtained easily. In this paper, we present an approach of incorporating unlabeled data for depen- dency parsing. First, all the sentences in unlabeled data are parsed by a dependency parser, which can provide state-of-the-art performance. We then ex- tract information on short dependency relations from the parsed data, because the performance for short dependencies is relatively higher than others. Fi- nally, we train another parser by using the informa- tion as features. The proposed method can be regarded as a semi- supervised learning method. Currently, most semi- 88 supervised methods seem to do well with artificially restricted labeled data, but they are unable to outper- form the best supervised baseline when more labeled data is added. In our experiments, we show that our approach significantly outperforms a state-of-the-art parser, which is trained on full labeled data. 2 Motivation …

Content truncated to 500 words