... iterationOnce at the beginning No supervised dataFigure 3: Dependence on the inclusion of the su-pervised training data (Czech)only) on the supervised training results, where the performance of the taggers ... all but one of the dimensions, andcompared their contribution (or rather, their losswhen compared to the best ”mix” overall). For English, we found that excluding the state-of -the- art-tagger ... al-ways uses the same tagger for tagging the unsu-pervised data. For both English and Czech, the selection oftaggers, the best combination and the best over-all setup has been optimized on the developmentdata...