Data and Methodology

All studies of register variation in parallel corpora are restricted to the types of texts available in the corpora, which, in turn, are restricted to texts that are often translated. According to Hasselgard (2010), “this precludes the study of many types of text, such as conversation, daily newspapers, and academic prose”. Our study focuses on three registers which are available in InterCorp: fiction, journalistic texts and spoken language. Unfortunately, journalistic texts are only represented by the PressEurope database, which does not contain any news reporting, and spoken language in InterCorp is far from spontaneous: our data come from the Proceedings from the European Parliament (Europarl) and from subtitles taken from the Open Subtitles Database. In addition, the data representing each register vary in size (the PressEurope subcorpus is the smallest and the subcorpus of Subtitles translated

Table 1 The size of the subcorpora (in text positions, TPs)

from English by far the largest; see Table 1) and come from different periods: the PressEurope texts in InterCorp cover the period between 2009 and 2014 and the Europarl texts date from 2007 to 2011. Our subcorpus of Fiction was created manually to include only books published after 1950 and to ensure that not more than two novels per author are included. As far as the subcorpora of Subtitles are concerned, since InterCorp only provides information about the year in which the original language version of the film was released (no information is provided for Czech as the target language), and since the number of Czech original films in InterCorp is very low, all Subtitle data available were included in the subcorpus.

This brings us to two more problems regarding comparability of the subcorpora: first, Czech as a small language is the source language of a much lower number of texts than English, and so subcorpora consisting of Czech target texts are always larger than subcorpora of Czech source texts. This applies to all subcorpora, as demonstrated in Table 1 above. Second, there is a problem with the concept of “original language”: while in our subcorpus of Fiction and of Subtitles Czech or English is always the language of the original, the same cannot be said about the PressEurope subcorpus; InterCorp does not provide information about the language of the original text for the texts included in the PressEurope database. Europarl (which is not annotated for the original language either) questions even the concept of the source language: as Gast and Levshina (2014, 377-378) argue, “until 2003 the texts were translated directly from the source languages into any of the target languages. From 2003 onwards ... all languages were first translated into English and then into the relevant target language”.^[1] What makes this even more complicated is the fact that a large proportion of the Europarl texts are not even annotated for the source language, which means, for example, that a potential subcorpus of Czech translations from English source texts only has 9,284 text positions (TPs) and pry does not occur in it. To obtain a sensible amount of Europarl data for analysis, following Gast and Levshina (2014),^[2] we resorted to a methodologically problematic solution not to differentiate between Czech as source and target language and to include even translations from other languages. This explains the single number in the last row in Table 1 above.

• [1] We hear that this practice is, however, abandoned if an interpreter between e.g. German andCzech is available; the translation then goes directly from German to Czech (Sarka Timarova, pers.comm.).
• [2] Gast and Levshina (2014, 379) argue that the “translations of the EUROPARL corpus are of avery high quality and certainly come close to that ideal”, namely the ideal of “near equivalence” ina translation corpus.