The automatic classification of abstract sentences into its main elements (background, objectives, methods, results, conclusions) is a key tool to support scientific database querying, to summarize relevant literature works and to assist in the writing of new abstracts. In this paper, we propose a novel deep learning approach based on a convolutional layer and a bidirectional gated recurrent unit to classify sentences of abstracts. First, the proposed neural network was tested on a publicly available repository containing 20 thousand abstracts from the biomedical domain. Competitive results were achieved, with weight-averaged Precision, Recall and F1-score values around 91%, and an area under the ROC curve (AUC) of 99%, which are higher when compared to a state-of-the-art neural network. Then, a crowdsourcing approach using gamification was adopted to create a new comprehensive set of 4111 classified sentences from the computer science domain, focused on social media abstracts. The results of applying the same deep learning modeling technique trained with 3287 (80%) of the available sentences were below the ones obtained for the larger biomedical dataset, with weight-averaged Precision, Recall and F1-score values between 73 and 76%, and an AUC of 91%. Considering the dataset dimension as a likely important factor for such performance decrease, a data augmentation approach was further applied. This involved the use of text mining to translate sentences of the computer science abstract corpus while retaining the same meaning. Such approach resulted in slight improvements (around 2 percentage points) for the weight-averaged Recall and F1-score values.