Two Case Studies on Translating Pronouns in a Deep Syntax Framework

International Joint Conference on Natural Language Processing, pages 1037–1041, Nagoya, Japan, 14-18 October 2013. Two Case Studies on Translating Pronouns in a Deep Syntax Framework Michal Nov´ak, Zdenˇek ˇZabokrtsk´y and Anna Nedoluzhko Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, CZ-11800 {mnovak,zabokrtsky,nedoluzko}@ufal.mff.cuni.cz Abstract We focus on improving the translation of the English pronoun it and English reflex- ive pronouns in an English-Czech syntax- based machine translation framework. Our evaluation both from intrinsic and ex- trinsic perspective shows that adding spe- cialized syntactic and coreference-related features leads to an improvement in trans- lation quality. 1 Introduction Machine Translation (MT) is an extremely broad task and can be decomposed along various di- rections. One of them lies in using specialized translation models (TMs) for certain types of lan- guage expressions. For instance, different types of named entities often receive specialized treat- ment in real translation systems. This paper deals with introducing specialized TMs for two types of pronouns: the pronoun it and reflexive pronouns. The models are integrated into an English-Czech syntax-based MT framework. Several works have previously focused on trans- lating pronouns. The linguistic study of Morin (2009) investigated the translation of pronouns, proper names and kinship terms from Indonesian into English. Onderkov´a (2010) has conducted a corpus-based research on possesive pronouns in Czech and English, focusing especially on their use with parts of the human body. From the perspective of MT, translating per- sonal pronouns from English to morphologically richer languages, such as French (Le Nagard and Koehn, 2010), German (Hardmeier and Federico, 2010) and Czech (Guillou, 2012) has recently aroused higher interest. In these languages, one usually has to ensure agreement in gender and number between the pronoun and its direct an- tecedent, which requires a coreference resolver to be involved. In this work, we make use of the English-to- Czech translation implemented within the Tec- toMT system (ˇZabokrtsk´y et al., 2008). In contrast to the phrase-based approach (Koehn et al., 2003), TectoMT performs a tree-to-tree machine transla- tion. An input English sentence is first analyzed into its deep-syntactic representation, which is subsequently transferred into Czech. The pipeline ends with generating a surface form of the Czech translation from its deep representation. The deep syntactic representation of a sentence in TectoMT follows the Prague tectogrammatics theory (Sgall, 1967; Sgall et al., 1986). It is a de- pendency tree whose nodes correspond to content words. Personal pronouns missing on the surface are reconstructed in special nodes. All nodes are assigned semantic roles and coreference relations are annotated. Originally, translation of both it and reflexive pronouns was treated by rules in TectoMT. The English deep representation of it was translated as to and a simple heuristics determined if it is be- ing expressed on the surface. Similarly, reflexives were always translated as se. This paper evalu- ates the translation quality reached using special- ized classifiers for these pronouns. Unlike the re- lated work on pronouns in MT, we focus on im- …

Content truncated to 500 words