Towards an integrated representation of multiple layers of linguistic annotation in multilingual corpora

Abstract

In the proposed talk we discuss the application of a set of computational text analysis techniques for the analysis of the linguistic features of translations. The goal of this analysis is to test two hypotheses about the specific properties of translations: Baker's hypothesis of normalization (Baker, 1995) and Toury's law of interference (Toury, 1995). The corpus we analyze consists of English and German original texts and translations of those texts into German and English, respectively. The analysis task is complex in a number of respects. First, a multi-level analysis (clause, phrases, words) has to be carried out; second, among the linguistic features selected for analysis are some rather abstract ones, ranging from functional-grammatical features, e.g., Subject, Adverbial of Time, etc, to semantic features, e.g., semantic roles, such as Agent, Goal, Locative, etc.; third, monolingual and contrastive analyses are involved. This places certain requirements on the computational techniques to be employed both regarding corpus encoding, linguistic annotation and information extraction. We show how a combination of commonly available techniques can fulfill these requirements to a large degree and point out their limitations for application to the research questions raised. These techniques range from document encoding (TEI, XML) over automatic corpus annotation (notably part-of-speech tagging; Brants, 2000) and semi-automatic annotation (O'Donnell, 1995) to query systems as implemented in e.g., the IMS Corpus Workbench (Christ, 1994), the MATE system (Mengel & Lezius, 2000) and the Gsearch system (Keller et al., 1999).