Mining Co-Occurrence Matrices for SO-PMI Paradigm Word Candidates

This paper is focused on one aspect of SOPMI, an unsupervised approach to sentiment vocabulary acquisition proposed by Turney (Turney and Littman, 2003). The method, originally applied and evaluated for English, is often used in bootstrapping sentiment lexicons for European languages where no such resources typically exist. In general, SO-PMI values are computed from word co-occurrence frequencies in the neighbourhoods of two small sets of paradigm words. The goal of this work is to investigate how lexeme selection affects the quality of obtained sentiment estimations. This has been achieved by comparing ad hoc random lexeme selection with two alternative heuristics, based on clustering and SVD decomposition of a word co-occurrence matrix, demonstrating superiority of the latter methods. The work can be also interpreted as sensitivity analysis on SO-PMI with regard to paradigm word selection. The experiments were carried out for Polish.