A Flexible Example Annotation Schema: Translation Corresponding Tree Representation

A Flexible Example Annotation Schema: Translation Corresponding Tree Representation Fai WONG, Dong Cheng HU, Yu Hang MAO Speech and Language Processing Research Center, Tsinghua University, 100084 Beijing huangh01@mails.tsinghua.edu.cn {hudc, myh-dau}@mail.tsinghua.edu.cn Ming Chui DONG Faculty of Science and Technology of University of Macao, PO Box 3001, Macao SAR dmc@inesc-macau.org.mo Abstract This paper presents work on the task of con- structing an example base1 from a given bi- lingual corpus based on the annotation schema of Translation Corresponding Tree (TCT). Each TCT describes a translation ex- ample (a pair of bilingual sentences). It repre- sents the syntactic structure of source language sentence, and more importantly is the facility to specify the correspondences be- tween string (both the source and target sen- tences) and the representation tree. Furthermore, syntax transformation clues are also encapsulated at each node in the TCT representation to capture the differentiation of grammatical structure between the source and target languages. With this annotation schema, translation examples are effectively represented and organized in the bilingual knowledge database that we need for the Por- tuguese to Chinese machine translation sys- tem. 1 Introduction The construction of bilingual knowledge base, in the development of example-based machine translation systems (Sato and Nagao, 1990), is vitally critical. In the translation process, the ap- plication of bilingual examples concerns with how examples are used to facilitate translation, which involves the factorization of an input sen- tence into the format of stored examples and the conversion of source texts into target texts in terms of the existing translations by referencing to the bilingual knowledge base. Theoretically speaking, examples can be achieved from bilin- 1 Or bilingual knowledge base, we use the two terms inter- changeably. gual corpus where the texts are aligned in senten- tial level, and technically, we need an example base for convenient storage and retrieval of ex- amples. The way of how the translation examples themselves are actually stored is closely related to the problem of searching for matches. In struc- tural example-based machine translation systems (Grishman, 1994; Meyers et al., 1998; Watanabe et al., 2000), examples in the knowledge base are normally annotated with their constituency (Kaji et al., 1992) or dependency structures (Matsu- moto et al., 1993; Aramaki et al., 2001; Al- Adhaileh et al., 2002), which allows the corre- sponding relations between source and target sen- tences to be established at the structural level. All of these approaches annotate examples by mean of a pair of analyzed structures, one for each lan- guage sentence, where the correspondences be- tween inter levels of source and target structures are explicitly linked. However, we found that these approaches require the bilingual examples that have ‘parallel’ translations or ‘close’ syntac- tic structures (Grishman, 1994), where the source sentence and target sentences have explicit corre- spondences in the sentences-pair. For example, in (Wu, 1995), the translation examples used for building the translation alignments are selected based on strict constraints. As a result, these ap- proaches indirectly limit their application in us- ing the translation examples …

Content truncated to 500 words