All parts are not created equal: SIAM-LSA

Previous research has shown the inadequacy of models for computing similarity that rely on any type of simple combination of features. Human similarity judgments are sensitive to the structure of the items being compared. For visual stimuli, the spatial arrangement of the items provides an obvious structure. For textual stimuli, however, the structure of the items must be inferred. Prior research on textual similarity has shown the dominant effect of relational features. We extend that research by looking at human judgments of the similarity of sentence pairs within the framework set out by Goldstone’s (1994) SIAM model, which calculates correspondences between objects and their features. We show that although the simple SIAM-based model fails to account well for the human judgments, a modified version which gives different weights to different semantic roles provides a strong match with human ratings.