Automatic alignment of Czech and English deep syntactic dependency trees

Automatic Alignment of Czech and English Deep Syntactic Dependency Trees⋆ David Mareˇcek, Zdenˇek ˇZabokrtsk´y, and V´aclav Nov´ak Institute of Formal and Applied Linguistics, Charles University in Prague, Czech Republic david.marecek@gmail.com, {zabokrtsky, novak}@ufal.mff.cuni.cz Abstract. In this paper, we focus on alignment of Czech and English tectogrammatical dependency trees. The alignment of deep syntactic de- pendency trees can be used for training transfer models for machine translation systems based on analysis-transfer-synthesis architecture. The results of our experiments show that shifting the alignment task from the word layer to the tectogrammatical layer both (a) increases the inter- annotator agreement on the task and (b) allows to construct a feature- based algorithm which uses sentence structure and which outperforms the GIZA++ aligner in terms of f-measure on aligned tectogrammatical node pairs. 1 Introduction Statistical machine translation requires a substantial amount of translation knowl- edge typically acquired from parallel corpora. For training a transfer on deeper syntactic layer it is feasible to use aligned deep syntactic trees. At first we will show the differences between alignment of sentences on the surface (word alignment) and alignment of their tectogrammatical representa- tions – deep syntactic dependency trees according to the specification of Prague Dependency Treebank 2.0 [1]. Tectogrammatical trees (t-trees for short, analo- gously t-layer and t-nodes) will be described in Section 2. The word alignment task is to find the most likely counterpart for every word in a sentence. It is questionable if we really need to find counterparts for all words, especially in the case of typologically different languages. For example, auxiliary words in one language differ in their functions and repertory from auxiliary words in another one. There is an example of English sentence and its Czech translation in Figure 1. The full arrows represent the obvious alignment pairs, whereas the correspon- dence expressed by the dashed arrows is not straightforward. For example, there is only one negation word No in the English sentence while in the Czech one, there is the negation in both ˇZ´adn´e and nebylo. The word nebylo can be trans- lated into English as wasn’t, but if the word dosud follows, the only possibility ⋆ The research reported on in this paper has been supported by the projects GAˇCR 405/06/0589, 1ET101120503, MSM0021620838, 1ET201120505, and LC 536. 12th EAMT conference, 22-23 September 2008, Hamburg, Germany 104 Fig. 1. Example of word alignment on the surface. is present perfect tense – has been. The word dosud has thus a relationship with the present perfect tense and should be linked besides yet also with has and been. This example illustrates the fact that the word-alignment of Czech- English sentence pairs is rather complex. [2] describe an experiment in which two annotators aligned manually 515 sentences from Czech-English corpus. The inter-annotator agreement of the simplest word alignment method (only one type of edge) reached 91%. In this paper we will be concerned with alignment of deep syntactic t-trees. On the t-layer the Czech and English sentence trees are more similar compared to the similarity …

Content truncated to 500 words