FUSION Dialogue System for Fuzzy Rule Matching

Dialogue systems frequently misinterpret user input containing vague, hedged, or perception-based expressions such as &#x201C;<italic>somewhat warm</italic>&#x201D; or &#x201C;<italic>not bright</italic>&#x201D;. Traditional sentence similarity measures rely on crisp semantic categories and struggle with the inherent fuzziness of natural language, leading to incorrect rule matching and dialogue breakdown. This paper presents a study comprising of two question-and-answer dialogue systems that integrate fuzzy semantic similarity measures based on Interval Type-2 fuzzy sets to address this limitation. The first system embeds a fuzzy similarity algorithm into a caf&#x00E9; customer-feedback scenario and was evaluated with 32 participants generating 288 responses. The second system incorporates algorithmic enhancements including a fuzzy negation operator and a cross-category influence factor, evaluated in a remote working context with 35 participants producing 630 utterances. Comparative evaluation against a WordNet-based baseline demonstrated that the fuzzy approach achieved substantially higher rule-matching accuracy (87.14% versus 69.05%), with particularly strong performance in subjective categories such as Temperature (96.88% versus 68.75%), Worth (90.63% versus 62.50%), and Speed (90.63% versus 62.50%). The study provides a complete implementation methodology including a five-phase linguistic preprocessing pipeline and an empirically validated fuzzy dictionary covering nine human perceptual categories. Results confirm that fuzzy semantic similarity offers a robust foundation for interpreting perception-driven language in dialogue systems.