GFIDF:Gradual Fusion Intent Detection Framework

Abstract In the field of multimodal intent detection research, researchers have worked on seamlessly integrating various information types to accurately recognize user intent, which is crucial for creating efficient dialog systems in complex real-world environments. Current approaches still have significant potential for improvement in exploring deep connections between modalities and extracting key semantic features from non-textual data. In this paper, we introduce GFIDF, a step-by-step fusion intention detection framework model, which first purifies and noise reduces the video modal information and audio modal information, and then takes the textual information as the starting point to align and fuse the video and audio modal features with the textual modal information effectively, respectively. Experiments conducted on the MIntRec dataset demonstrate the effectiveness of the model in multimodal feature fusion, thus validating its research value.