Linear separability in superordinate natural language concepts.

Two experiments are reported in which linear separability was investigated in superordinate natural language concept pairs (e.g., toiletry-sewing gear). Representations of the exemplars of semantically related concept pairs were derived in two to five dimensions using multidimensional scaling (MDS) of similarities based on possession of the concept features. Next, category membership, obtained from an exemplar generation study (in Experiment 1) and from a forced-choice classification task (in Experiment 2) was predicted from the coordinates of the MDS representation using log linear analysis. The results showed that all natural kind concept pairs were perfectly linearly separable, whereas artifact concept pairs showed several violations. Clear linear separability of natural language concept pairs is in line with independent cue models. The violations in the artifact pairs, however, yield clear evidence against the independent cue models.