Adjective Intensity Ordering by Representing Word Definitions as a System of Linear Equations
Stefan E. Eng, Jennifer Tan, Markus R. Iseli · 2018
While speakers may disagree or have difficulty deciding the ordering of semantically similar adjectives such as icy and frosty, arranging adjectives on a linear scale is an intuitive task for most speakers of any language. Currently, there is no standard method or database to order gradable adjectives that share a common attribute, such as temperature or speed. Modeling how adjectives relate to one another semantically is an important aspect of language understanding. Applications of adjective intensity ordering include detecting scale differences in reviews and measuring similarity of larger segments of text. This paper discusses a novel approach that solves a system of linear equations constructed only from word dictionary definitions to compare the intensity of similar adjectives. We used Krippendorff's alpha to measure the agreement between a human standard and our algorithms' results on an ordinal scale and attained a score of 0.927 for temperature-, 0.652 for speed-, and 0.673 for happiness-related adjectives. As a reference, Krippendorff inter-rater agreement was 0.927 for temperature, 0.816 for speed-related adjectives, and 0.849 for happiness-related adjectives.