With regard to grammar checking, the number of unnecessary interventions is obviously considerably higher than the correct ones for the corrected corpora.
Anomalies are detected using heuristics which we developed after analyzing our corpora.
As extensive controlled annotated corpora were not still available at the time of the experiments, resources have been manually derived for them.
Table 2 gives some statistics about the corpora, the participants, and the scores of the best performing systems.
Recent approaches to discourse processing, on the other hand, that allow robust parsing require extensive corpora investigations.
A third limitation is that the computation of all predictor variables was based on adult corpora.
Thus, all predictor variables were based on adult corpora for uniformity across variables.
However, the overall rates of orthography influence are low in all three corpora.
