An Experimental Study Comparing Human with Corpus-derived Associations

Collocations often comprise two components where one dominates the other in the sense that knowing the dominant one makes it almost inevitable to think of the other one as well. In contrast, knowing the non-dominant component does not necessarily evoke such a strong preference regarding the missing component. In the combination Pyrrhic victory, for example, the first component almost “demands” the second whereas the second does not suggest the first as strongly. Moreover, a range of other words that can go with victory come to mind, for example narrow, landslide or decisive. The topic of this thesis is to examine to what extent corpus-derived association measures are able to capture these asymmetric relations that can occur in collocations. To this end, four asymmetric measures that can express different strengths of association between two components of a collocation were defined. The new measures are based on well-established association measures and were applied to a large data set. In order to evaluate their performance against empirical data, an experiment with human subjects was carried out. The measures are shown to be capable of predicting the direction of association within collocations. Accuracy varies between 76% and 86% for the different measures.