Using the shape recovery method to evaluate indexing techniques

Text representation, central to information processing, must be descriptive and discriminative. Although some of the many techniques to construct document representations may outperform others for certain tasks, no one is consistently better than others. Representations are still problematic. Evaluation techniques are needed to penetrate foundational questions about term behavior in representation. A study that applies the shape recovery analysis method is reported here as an evaluative tool to compare different indexing schemes. Three weight coefficients are used to rank indexing terms and are compared to the documentsʹ full text. Two of the weight coefficients are novel and the third relies on the chi-squared distribution. Multidimensional scaling reduces the dimensional space of the document surrogates into a two-dimensional Cartesian space. Ten concentric circles evenly separated at 10% intervals of relevant data points starting at the centroid are used to construct a precision–recall curve. ANOVA is used for a straightforward computation of the 4 × 11 matrix of test data to see whether the four treatments yield the same P-R result. A post hoc HSD Tukey multiple comparisons test among pairwise treatments is also used to discover homogeneous groups. The findings show the value of the methodology to study term weighting schemes, and their descriptiveness and discriminative power, as well as the potential strength of the novel coefficients introduced.