ZIPPER: The holistic spell checker

Abstract

This article discusses a novel problem with a novel solution. The thesis of this work is to perform document and text editing. We look specifically at providing a document editor with a tool for text editing and that tool is holistic spell checker. We propose ZIPPER, a holistic spell checker, that finds out user misspell words all at once. Our approach to this approximated problem is probabilistic. We use Markovian tree that exploits the dependencies amongst characters in a word. For computation over Markovian tree, we use a set of probabilistic and information theory metrics. For link quantification, we use information theory metric which is pointwise mutual information. Probabilistically, we use belief propagation using message passing paradigm for node quantification. To create a suggestion list for each misspelled word, we decompose Markovian tree into clique tree. For computation over clique tree, where each node is a complete word, we use information theory metrics like: entropy for computing the value of a complete word, and mutual information for computing how much value there is between two words.