Automatic detection of emotions in textual data masses provides priceless opportunities for researchers and also it is inevitable for practitioners. The unfavorable factors involved in text data cause ambiguity and adversely affect the performances of emotion classifiers. Although deep learning approaches spark off significantly successive results, the obtained performances of the classifiers in the literature are commonly evaluated as the overall accuracy. This incomplete evaluation ignores inner class performance and overall accuracy can behave as a hopeful evaluator. In this study, we employed deep learning and meta-heuristic optimization methods together in order to resolve the ambiguity issue. Moreover, the decision mechanism of a conventional deep learning model is equipped with optimal emotion vectors obtained by optimization processes for each emotion class. Experimental results show that the proposed approach improves the inner class performance by maintaining the overall accuracy scores.