Automatic Enhancement of LTAG Treebank

Proceedings of Recent Advances in Natural Language Processing, pages 733–739, Hissar, Bulgaria, 7-13 September 2013. Automatic Enhancement of LTAG Treebanks Farzaneh Zarei, Ali Basirat*, Heshaam Faili and Maryam Sadat Mirian School of Electrical and Computer Engineering College of Engineering University of Tehran, Tehran, Iran {zareeifarzaneh,hfaili,mmirian}@ut.ac.ir *a.basirat@srbiau.ac.ir Abstract The Treebanks as the sets of syntactically an- notated sentences, are the most widely used language resource in the application of Natural Language Processing. The occurrence of er- rors in the automatically created Treebanks is one of the main obstacles limiting the using of these resources in the real world applications. This paper aims to introduce an statistical me- thod for diminishing the amount of errors oc- curred in a specific English LTAG-Treebank proposed in Basirat and Faili (2013). The problem has been formulated as a classifica- tion problem and has been tackled by using several classifiers. The experiments show that by using this approach, about 95% of the er- rors could be detected and more than 77% of them could successfully be corrected in the case of using Adaboost classifier. In addition, it has been shown that the new treebank could reach a high of 76% F-measure which is 8% higher than the original treebank. 1 Introduction Treebanks, as special corpora annotated with syntactic structures, play a crucial role in the re- cent success of natural language processing ap- plications like speech recognition, spoken lan- guage systems (Xue et al, 2005), parsing (Mirro- shandel et al, 2012), and machine translation (Kotze et al, 2012). Regarding the development methods of the tree- banks, generally, they can be placed in either manually crafted or automatically extracted tree- banks. Due to the large number of sentences, the manual creation of the treebanks can be very ex- pensive and time consuming. For instance, Penn English treebank as one of the outstanding handmade ones took eight years (1989-1996) to be completed. The difficulties, raised in the ma- nual creation of Treebanks, led the researchers to use automatic and semi-automatic methods of treebank development methods. On the other hand, the automatically extracted Treebanks are not as accurate as manual versions. In fact, these resources mostly suffer from the occurrence of error in the annotated sentences that in turn re- duces the applicability of these resources in the real world applications. A large number of researchers tried to improve the quality of the automatically extracted Tree- banks in order to increase the applicability level of these resources in the NLP tasks (Xue et al, 2005). For instance Dickinson and Meurers (2003) proposed an n-gram based approach for detecting Part-of-Speech errors in Penn English Treebank. In other works, Agarwal et al. (2012) proposed a hybrid approach to improve the mechanism of error detection introduced by Ambati (2011) for detecting the errors in a dependency treebank. Ule and Simov (2003) also could find unex- pected tree productions by using a method called Directed Treebank Refinement (DTR). In this paper, we try to correct the errors oc- curring in the treebank automatically …

Content truncated to 500 words