Sentiment analysis datasets usually have class imbalance, with far more positive reviews than negative ones, so looking only at overall accuracy means a model that predicts everything as positive can still get a very high score.