Identifying Collocations to Measure Compositionality: Shared Task System Description

This paper describes three systems from the University of Minnesota, Duluth that participated in the DiSCo 2011 shared task that evaluated distributional methods of measuring semantic compositionality. All three systems approached this as a problem of collocation identification, where strong collocates are assumed to be minimally compositional. duluth1 relies on the t-score, whereas duluth-2 and duluth-3 rely on Pointwise Mutual Information (pmi). duluth-1 was the top ranked system overall in coarse–grained scoring, which was a 3-way category assignment where pairs were assigned values of high, medium, or low compositionality.