Generation of Word Profiles on the basis of a large and balanced German corpus

Electronic corpora have been used in lexicography and the domain of language learning for more than two decades (cf. Braun et al. 2006, Sinclair 1991). Traditionally, computer platforms exploiting these corpora were based on concordances that present a word in its different contexts. However, concordances hit their limits for very large corpora where the result sets are generally too large for manual evaluation. To answer questions like 'which attributive adjectives are used for the noun book' or 'is the adjective groundbreaking more typical for book than pioneering', would require one to look at several thousand concordance lines, a quite impracticable task to do by hand. Likewise, the exclusive use of concordance lines in an attempt to answer a question like 'which objects does a verb like hit typically take' would be unsuitable, since one would not only have to find all the different objects of hit but it would also be necessary to discard all the false positives. These types of questions involve counting of co-occurrences, and, if they are linguistically motivated, collocations. The cases above are examples for collocations of a certain syntactic type, i.e. adjective-noun and verbobject collocations. The importance of describing collocations has long been acknowledged both for language learning (e.g. Hausmann 1984) as well as for lexicographic purposes (e.g. Harris 1968). Church & Hanks (1989) were the first to show that lexical statistics are useful to summarize concordance data by presenting a list of the statistically most salient collocates. More recently, databases have been built for large corpora that make use of this abstraction of concordance lines. Examples are Lexiview, an interactive platform for German supporting the manual work of the lexicographer (Evert et al. 2004), or the Sketch Engine (Kilgarriff 2004) that produces so called 'word-sketches' for languages as different as Czech, Italian or Chinese. Both approaches provide lists of the statistically most salient collocates for each grammatical relation in which the word participates.