Combining coregularization and consensus-based self-training for multilingual text categorization