FUZZY-BAYESIAN-PARTNERSHIP TOWARDS THE INFERENCE OF CHEREMES IN HUMAN SIGN LANGUAGES

摘要

We characterize all the distinct ways one chereme (hand-shape) of the Spanish Sign Language may be performed (the cheretypes). The solution is based on maximal sub-relations of similarity among probability distributions, type 2 fuzzy sets, and Bayesian inference. Crisp evidence consists of the normalized joint angle values coming from sensor devices such as those mounted on instrumented gloves. We work with fuzzy-valued evidence describing cheremes and cheretypes as type 2 fuzzy subsets of the whole collection of sensors in order to incorporate linguistic evidence into the system (an example of an input fuzzy grade is: "medium flexed"). We integrate the multiple expert information identifying the "brute" cheretypes with maximal sub-relations of similarity. Using a fuzzy Bayesian scheme we first estimate the likelihood of crisp evidence from the likelihoods of linguistic evidence and then we compute the posterior probabilities allowing the selection of the most probable and most relevant cheremes.