Identifying Phrasemes via Interlingual Association Measures - A Data-driven Approach on Dependency-parsed and Word-aligned Parallel Corpora

It has been understood for a long time that the semantic content of a combination of two or more words often cannot be derived from the semantics of the single words, but that the use of one particular word imposes restrictions upon others (Firth 1957; Evert 2004, 15–17). The semantics is then either determined by the ruling word, e.g., in the case of light verb constructions (attention entails pay in pay attention), or by the entity of all participating words, e.g., in the case of idiomatic expressions or set phrases (so to speak). Many names have been given to this phenomenon, each of which looking at it from a slightly different perspective: collocations, multiword expressions, phrasemes, idioms or formulaic sequences, to name just a few. The term lexical function (Wanner 1996; Mel’čuk 1998) stresses the aspect of one word being the value returned by a function applied to another word. That lack of flexibility of the determined word, which does not contribute much – if anything – to the meaning of the composed expressions is what we make use of in the approach described in this paper. In our work, we address the issue of phraseme identification in one language by searching for corresponding syntactic structures in parallel, word-aligned corpora. We exemplify our approach by retrieving and ranking support verb constructions that consist of a verb and its direct object. The support verb’s nature allows for correspondences that can be regarded as translations only in the context of the whole construction. A suitable translation of the support verb construction pay attention into German is Aufmerksamkeit schenken ‘attention’ + ‘give as a present’/‘make a gift’. While attention and Aufmerksamkeit embody the same semantic concept, pay and schenken can hardly be seen as good translations except for this particular case – and only in conjunction with their direct objects. This paper is structured as follows: In section 2, we give an overview of statistical association measures and their motivation. Section 3 explains the design of our corpus including the choice of corpus material and the annotation and alignment tasks that are required to allow for complex corpus queries such as the ones we use for