Calibrating for Specific Domains

Abstract

Eigenspace-based models were shown to exhibit greater effectiveness than their simple vector space counterparts in settings that benefit from fuzziness (such as information retrieval or recommender systems). In settings that require precision in representation structure (such as in essay scoring or for conceptual relationship mining), however, improved means to predict model behaviour from parameter settings could ease applicability and increase efficiency by reducing tuning times.

This chapter reports experiences and experiment results from a systematic investigation of tuning parameters, their potential settings, and interdependencies between them. This includes studying the influence of sanitising operations, sampling, dimensionality changes, and degrees of specialisation. Trends indicate that the smaller the corpus, the more domain-specific documents are required. Moreover, recommendations for vocabulary filtering can be derived, dependent on the size of the corpus.