Text-based emotion detection (ED) is a developing area of study in text classification and Natural Language Processing (NLP). People use social platforms such as Twitter, Facebook, Instagram, etc., to express their views or emotions with others. The analysis of emotions on such platforms helps us to understand the user’s feelings about every little thing surrounding them. However, the performance and efficiency of emotion analysis has been affected by the challenges of NLP. Recent studies show that deep learning (DL) models performed well and provided promising results as compared to Machine Learning (ML) models for various NLP-related tasks. In DL models, word embedding is used to generate input feature vectors. At present, different word embedding models have been developed to represent words in the vector form. The accuracy of DL models depends upon the type of word embedding used. Thus, choosing the right word embedding technique can affect the performance of emotion classification models. Hence, this paper aims to compare different word embeddings, namely FastText, Glove, Word2Vec, and BERT using various DL models for ED on the Amazon dataset. The main objective of our work is to identify the most effective combination of classification algorithms and word embedding models for emotion analysis feature engineering tasks and emotion classification, respectively. We observed that BERT outperforms the other models. BERT used with BiLSTM, CNN, and Bi-GRU performed well as compared to other combinations. Our result also shows that word embeddings have a significant effect on the performance of emotion classifiers.