High Quality Dependency Selection from Automatic Parses

International Joint Conference on Natural Language Processing, pages 947–951, Nagoya, Japan, 14-18 October 2013. High Quality Dependency Selection from Automatic Parses Gongye Jin, Daisuke Kawahara, Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-Honmachi, Sakyo-ku, Kyoto, 606-8501, Japan jin@nlp.ist.i.kyoto-u.ac.jp {dk,kuro}@i.kyoto-u.ac.jp Abstract Many NLP tasks such as question answer- ing and knowledge acquisition are tightly dependent on dependency parsing. Depen- dency parsing accuracy is always decisive for the performance of subsequent tasks. Therefore, reducing dependency parsing errors or selecting high quality dependen- cies is a primary issue. In this paper, we present a supervised approach for au- tomatically selecting high quality depen- dencies from automatic parses. Experi- mental results on three different languages show that our approach can effectively se- lect high quality dependencies from the re- sult analyzed by a dependency parser. 1 Introduction Knowledge acquisition from a large corpus has been actively studied recently. Knowledge is often acquired from the fundamental analysis. In partic- ular, dependency parsing has been used for some tasks like case frame compilation (Kawahara and Kurohashi, 2006), relation extraction (Saeger et al., 2011) and paraphrase acquisition (Hashimoto et al., 2011). For these tasks, the accuracy of de- pendency parsing is vital. Although the accuracy of state-of-the-art dependency parsers for some languages like English or Japanese is over 90%, it is still not high enough to acquire accurate knowl- edge, not to mention those difficult-to-analyze lan- guages like Chinese and Arabic. Instead of using all the automatic parses, it is possible to use only high quality dependencies for knowledge acquisition. In this paper, we present a supervised approach for selecting high quality dependencies from automatic dependency parses. This method considers linguistic features that are related to the difficulty of dependency parsing. The experimental results on English, Chinese and Japanese show that our proposed method can se- lect dependencies of higher quality than baseline methods for all the languages. 2 Related Word There have been a few approaches devoted to au- tomatic selection of high quality parses or depen- dencies. According to selection algorithms, they can be categorized into supervised and unsuper- vised. Supervised methods mainly focus on the con- struction of a machine learning classifier to pre- dict the reliability of parses or dependencies based on various kinds of features both on syntactic and semantic level. Yates et al. (2006) created WOODWARD which is a Web-based semantic fil- tering system. Kawahara and Uchimoto (2008) built a binary classifier that classifies each parse of a sentence as reliable or not. Among super- vised methods, ensemble approaches were also proposed. Reichart and Rappoport (2007) de- tected parse quality by a Sample Ensemble Parse Assessment (SEPA) algorithm. Another similar approach proposed by Sagae and Tsujii (2007) also selected high quality parses by computing the level of agreement on different parser out- puts. Iwatate (2012) applied a tournament model on Japanese dependency parsing and then selected reliable dependencies by using SVM output. The work most related to ours is the work of Yu et al. (2008). They proposed a framework …

Content truncated to 500 words