Adjective Deletion for Linguistic Steganography and Secret Sharing

This paper describes two methods for checking the acceptability of adjective deletion in noun phrases. The first method uses the Google n-gram corpus to check the fluency of the remaining context after an adjective is removed. The second method trains an SVM model using n-gram counts and other measures to classify deletable and undeletable adjectives in context. Both methods are evaluated against human judgements of sentence naturalness. The application motivating our interest in adjective deletion is data hiding, in particular linguistic steganography. We demonstrate the proposed adjective deletion technique can be integrated into an existing stegosystem, and in addition we propose a novel secret sharing scheme based on adjective deletion.