Looking for a needle in a haystack: Semi-automatic creation of a Latvian multi-word dictionary from small monolingual corpora

Multiword expressions (MWEs) are an indispensable part of almost any dictionary. However, the identification of missing MWEs that have recently appeared in a language is not a simple task. In this paper we describe automated methods for MWE identification in a rather small Latvian text corpora. We propose starting with the application of statistical measures to identify a wide range of MWEs and then applying linguistically motivated filters to clean the list of initially extracted MWE candidates. We show that for morphologically rich languages, such as Latvian, in cases with a small amount of language data better results can be achieved with lemmatized data. We also demonstrate that in the case of a small general domain (balanced) corpus, automatic methods can be used to find good MWE candidates - terminological units, named entities and some lexicalized phrases. However, finding idiomatic expressions in small, general domain corpora is looking for a needle in a haystack: only a larger or more expressive corpus can help in the identification process.