Text Classification Model Using Weighted Word Embeddings

The patent applies different weights to different words based on their informativeness. Specifically, it uses inverse document frequency (IDF) to calculate weights where common words receive lower weights and rare/informative words receive higher weights. This local differentiation resolves the contradiction by maintaining simplicity through a systematic weighting approach while improving accuracy through selective emphasis on meaningful words.

The patent changes the parameter of word weighting from uniform (equal weight) to variable (different weights based on IDF). By introducing the weight parameter w_i = log(N/n_i) where N is total documents and n_i is document frequency of word i, the model transforms from a simple average to a weighted average that preserves classification accuracy while maintaining computational efficiency.

The patent extracts and removes the harmful effect of common words by assigning them weights close to zero through the IDF calculation. Words like 'the', 'a', 'is' that appear in nearly all documents receive weights of log(N/N) = log(1) = 0, effectively taking them out of the classification calculation while preserving computational efficiency through the simple logarithmic formula.

The patent converts the harmful presence of common words into a benefit by using their high document frequency as the basis for assigning them low weights. The same property that makes them noisy (appearing everywhere) becomes the mechanism for their suppression (log(N/n_i) approaches 0 when n_i approaches N), turning the noise problem into a solution.