Semimage: HSV-Based Semantic Image Encoding for Disentangled Text Representation
Researchers have developed SemImage, a novel method that transforms text documents into 2D semantic images for processing by convolutional neural networks. Each word is mapped to a pixel with disentangled HSV color values representing topic, sentiment, and intensity. This approach, which includes dynamic boundary rows between sentences to highlight semantic shifts, has shown competitive or superior performance in document classification tasks compared to established models like BERT. AI
IMPACT Offers a new visual representation for text that can be processed by CNNs, potentially improving document analysis and interpretability.