Traditional models suffer from feature sparsity， feature loss and incomplete comment feature extraction problems due to the imbalance of comment length. This paper proposes an emotional analysis approach based on dynamic word-sentence features and self-attention （DWSF-SA）， to alleviate the incomplete extraction problem caused by the imbalance of text size under batch training. DWSF-SA first follows pre-training on dynamic feature embedding， then employs sentence vectors to complete the less parts and represents the truncated parts by fixed length. Moreover， DWSF-SA also introduces a self-attention mechanism to dynamically integrate the word-sentence fusion features， and makes optimization on the weight parameters to accelerate the computation and training. The ablation and comparison experiments on publicly available datasets demonstrate that the proposed DWSF-SA outperforms traditional approaches in accuracy metrics.