This study proposes an improved Multi-label K-Nearest Neighbors (MLkNN) classifier for multi-label emotion classification of short texts, particularly tweets. The improvement involves considering in-sentence features, adjacent sentence features, and full-text features to iteratively refine emotion classifications. Experiments on a Twitter corpus compare the base MLkNN, a sample-based version (S-MLkNN), and a label-based version (L-MLkNN). Results demonstrate that the improved L-MLkNN achieves superior accuracy and recall, especially with K=8 and α=0.7, showcasing the effectiveness of the iterative correction method.
