Corpus-based Acquisition and Analysis of Support Verb Constructions

This thesis deals with Support Verb Constructions (SVCs) and their automatic acquisiton. SVCs are verbal strucures, consisting of a verb and a noun, which form a unit in both syntactic and semantic aspects. As SVCs are hard to interpret on both counts, they are especially challenging for natural language processing. We test the possibilities of the acquisiton of SVCs by means of corpusbased methods with few linguistic resources. In particular, we investigate the phenomenon in Portuguese. The acquisition is carried out in a two-stage approach. First, we extract SVCs using a bilingual parallel corpus. Starting from a list of Portuguese full verbs which approximately correspond to the meaning of an SVC, we use the alignment information to retrieve Portuguese expressions which are semantically appropriate SVCs. In this context, the parallel language acts as a ‘pivot’ to connect the Portuguese full verb and SVCs. In the next step, we analyse the possibilities to refine the retrieved expressions. It turns out that it is difficult to use information about the support verb’s arguments to do such a filtering. Instead, we calculate association measures (e.g. pointwise mutual information) and compile a ranking. This second step, thus, is conducted on the monolingual level. The experiments show that the presented approach works very well: we retrieve semantically appropriate SVCs and achieve a maximum precision of 91% and a maximum recall of 86% in two different settings. However, the applicability of the approach depends on the contextual diversity of the initial full verb. Heterogeneity complicates the acquisition of high quality SVCs.