Exploring Treebank Transformations in Dependency Parsing

International Conference RANLP 2009 - Borovets, Bulgaria, pages 33–38 Exploring Treebank Transformations in Dependency Parsing Kepa Bengoetxea IXA NLP Group Technical School of Engineering, Bilbao University of the Basque Country Plaza La Casilla 3, 48012, Bilbao kepa.bengoetxea@ehu.es Koldo Gojenola IXA NLP Group Technical School of Engineering, Bilbao University of the Basque Country Plaza La Casilla 3, 48012, Bilbao koldo.gojenola@ehu.es Abstract This paper presents a set of experiments performed on parsing the Basque Dependency Treebank. We have concentrated on treebank transformations, maintaining the same basic parsing algorithm across the experiments. The experiments can be classified in two groups: 1) feature optimization, which is important mainly due to the fact that Basque is an agglutinative language, with a rich set of morphosyntactic features attached to each word, 2) graph transformations, ranging from language independent methods, such as projectivization, to language specific approaches, as coordination and subordinated sentences, where syntactic properties of Basque have been used to reshape the dependency trees used for training the system. The transformations have been tested independently and also in combination, showing that their order of application is relevant. The experiments were performed using a freely available state of the art data-driven dependency parser [11]. Keywords Dependency parsing, treebank parsing, agglutinative language. 1 Introduction This work presents several experiments performed on dependency parsing of the Basque Dependency Treebank (BDT) [1]. Several syntactic analyzers based on dependencies have been developed, with proposals ranging from systems that directly construct dependency structures [9] to other systems based on the more traditional constituency structures that allow the extraction of dependencies [2]. The present work has been developed in the context of dependency parsing exemplified by the CoNLL1 shared task on dependency parsing in years 2006 and 2007 [12], where several systems had to compete analyzing data from a typologically varied range of 11 languages. The treebanks for all languages were standardized using a previously agreed CONLL-X format (see Figure 1). BDT was one of the evaluated treebanks, which will allow us to make a direct comparison of results. Many works on treebank parsing have dedicated an effort to the task of pre-processing training trees [4, 13]. This paper extends these works, applying treebank 1 CoNLL: Computational Natural Language Learning. transformations [7, 10] to a morphologically rich, agglutinative language. The rest of the paper is organized as follows. Section 2 presents the main resources used in this work, including the BDT and a data-driven open source parser. Section 3 presents the different proposals for Treebank transformation that have been devised in order to improve the parser’s accuracy. Next, section 4 will evaluate the results of each transformation. Section 5 examines related work, and the last section outlines the main conclusions. 2 Resources This section will describe the main elements that have been used in the experiments. First, subsection 2.1 will present the Basque Treebank data, while subsection 2.2 will describe the main characteristics of Maltparser, a state of the art and data-driven dependency parser. 2.1 The Basque Dependency Treebank BDT [2] can be …

Content truncated to 500 words