Discovering Continuous Multi-word Expressions in Czech

Multi-word expressions frequently cause incorrect annotations in corpora, since they often contain foreign words or syntactic anomalies. In case of foreign material, the annotation quality depends on whether the correct language of the sequence is detected. In case of inter-lingual homographs, this problem becomes difficult. In the previous work, we created a dataset of Czech continuous multi-word expressions (MWEs). The candidates were discovered automatically from Czech web corpus considering their orthographic variability. The candidates were classified and annotated manually. Afterwards, the dataset was extended automatically by generating all word forms of those MWEs that were annotated as nouns. In this work, we used the dataset as positive examples, we filtered out negative examples from the MWE candidates. We trained a classifier with mean accuracy 92.7%. We have shown that the combined approach slightly outperforms approaches concerning only association measures mainly on MWEs containing inter-lingual homographs and out-of-vocabulary words. The discovery methods can be applied to other languages which encounter orthographic variability in web corpora.