PhD thesis - School of Informatics - University of Edinburgh

Chapter 4. System Extension to a New Language 104
that they are less repetitive than those contained in the German articles. However,
overall TTRs are 0.6 points lower for French than for German which means that the
remaining vocabulary in the French articles is somewhat less heterogeneous than in
the German data. This is due to the nature of the two languages. In German, lexical
variety is higher due to compounding and case inflection for articles, adjectives and
nouns.
French internet data
Token f Token f
e 49 web 21
Google 44 spam 18
internet 35 mail 18
Microsoft 33 Firefox 16
mails 32 Windows 14
Table 4.2: Ten most frequent (f) English inclusions in the French data.
The ten most frequent English inclusions found in the French data set are listed
in Table 4.2. The majority of them are either common or proper nouns. The most
frequent English token is actually the single-character token e as occurring in e-mail.
Less frequent pseudo-anglicisms like le forcing in the sense of putting pressure on
someone or le parking referring to a car park are also contained in the French data.
Although such lexical items are either non-existent or scarcely used by native English
speakers, they are also annotated as English inclusions.
4.3 System Module Conversion to French
The extended system architecture of the English inclusion classifier consists of sev-
eral pre-processing steps, followed by a lexicon module, a search engine module and
a post-processing module. Converting the search engine module to a new language
required little computational cost and time. Conversely, the pre- and post-processing
as well as the lexicon modules necessitated some language knowledge resources or