Annotation of Multi-Word Expressions in Czech Texts

Multi-word expressions (MWEs) are difficult to define and also difficult to annotate. Some of them cause serious errors in the traditional annotation pipeline tokenization – morphological analysis – morphological disambiguation. Many cases of incorrect annotation in Czech corpora are known. To narrow the research topic, we focus only in fixed MWEs – those with fixed word order and no ellidable components. In this paper, we propose a corpus-based method that reveals fixed MWE candidates. From the web-based corpus of Czech, we extracted 25,091 expressions, 2,140 of them were identified as MWEs, 332 as probable MWEs, and 174 of them can be either MWEs or one single word. Our method is based on corpus data observation that indicates that people are unsure when writing a MWE whether it is one word, a word with dashes, or several words. The result is a list of MWE candidates and also an application that classifies the input as MWE, probable MWE, or non-MWE.