Confidence-Aware Semantic Group Classification for Adjective Substitution in Linguistic Steganography

Linguistic steganography conceals information by replacing words in a natural language text. Existing methods based on masked language models (MLMs) or synonym substitution suffer from semantic ambiguity, for example an adjective can belong to many semantic groups depending on context, leading to unnatural substitutions, extraction failures, and multiround inconsistency. This paper proposes a Confidence-Aware Semantic Group Classification (CASGC) framework that validates the semantic group for an adjective before any embedding is decided. The framework first extracts adjective-noun pairs through dependency parsing after giving the sentence to the model containing adjective(s), each adjective is then classified into one of the twelve predefined semantic groups using SentenceBERT similarity, calculate the margin between the top two group prediction scores to get the confidence status, and applies a threetier decision gate (Confident, Borderline, Ambiguous) to decide whether the adjective is suitable for steganographic substitution. The experimental evaluation across four context representation strategies and six polysemantic adjective pairs shows that fullsentence context adjectives the highest classification accuracy of 83.33%, representing a 16.7% improvement over the refined baseline. The results suggest that semantic confidence analysis can be a promising prerequisite for reliable linguistic steganography based on adjectives.