Assigning importance scores to individual words or subwords in text, allowing the model to emphasize semantically significant terms in its representation.