PhD thesis - School of Informatics - University of Edinburgh
PhD thesis - School of Informatics - University of Edinburgh
PhD thesis - School of Informatics - University of Edinburgh
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Chapter 4. System Extension to a New Language 114<br />
development and test sets versus 0.29 and 0.25 in German development and test sets,<br />
see Table 4.1). This means that the classifier is already less likely to miss inclusions<br />
which minimises the effect <strong>of</strong> consistency checking for French.<br />
Test set Development set<br />
Method Acc P R F Acc P R F<br />
French data<br />
FS 98.10% 88.59% 84.11% 86.29 98.45% 91.60% 84.08% 87.68<br />
FS+CC 98.08% 88.35% 84.30% 86.28 98.49% 91.39% 85.09% 88.13<br />
German data<br />
FS 97.93% 92.13% 75.82% 83.18 98.07% 93.48% 73.31% 82.17<br />
FS+CC 98.13% 91.58% 78.92% 84.78 98.25% 92.75% 77.37% 84.37<br />
Table 4.5: Evaluation <strong>of</strong> the full system (FS) with optional consistency checking (CC).<br />
4.5 Chapter Summary<br />
This chapter described how the English inclusion classifier was successfully converted<br />
to a new language, French. The extended system is able to process either German or<br />
French text for identifying English inclusions. The system is a pipeline made up <strong>of</strong> sev-<br />
eral modules, including pre-processing, a lexicon, a search-engine, a post-processing<br />
and a document consistency checking module. The extension <strong>of</strong> the core system was<br />
carried out in only one person week and resulted in a system performance <strong>of</strong> 71.59<br />
points in F-score on the French development data. A further week was spent on im-<br />
plementing the post-processing module which boosted the F-score to 87.68 points. A<br />
third week was required to select external language resources plus collect and annotate<br />
French evaluation data in the domain <strong>of</strong> internet and telecoms.<br />
The performance drop between the French development set and the unseen test sets<br />
is relatively small (1.85 in F-score) which means that the system does not seriously<br />
over- or undergenerate for this domain but results in an equally high performance on<br />
new data. This chapter also demonstrated that the English inclusion classifier is easily<br />
extendable to a new language in a relative short period <strong>of</strong> time and without having to