Sentiment analysis (SA) is an automated process that is increasingly being applied on software engineering (SE) text to study developers’ sentiments and to improve productivity. Prior studies have reported varying performance for software engineering datasets depending on the techniques employed. In this paper, we propose deep-learning approach to perform sentiment analysis to identify the sentiment polarities of Jira  dataset. We developed and compared the performance of stacked CNN, stacked LSTM and stacked BiLSTM classifiers and found that the stacked BiLSTM classifier outperformed the other classifiers developed on the above-mentioned dataset. Additionally, this study investigated the impact of domain-customization on the performance of the classifiers by employing SE-specific word embeddings learnt from StackOverflow posts and compared with the pre-trained open-domain word embeddings learnt from google news posts. The classifiers employing Google News (GN) embeddings outperformed the SE-customized based classifiers. The better performance of GN embedding classifiers is due to the large generic training corpus of the GN word embedding model which identifies the sentiments expressed accurately when compared to SE-specific word embedding.