Today, human society suffers from the control of social media in all aspects of life. Despite the facilities provided by social media, it has also caused a threat to society economically, socially, and politically by publishing fake news that raises political and societal divisions between members of society and each other on the one hand and between members of society and governments on the other. Therefore, in recent years, there has been an active effort to build systems that could distinguish between news types (real and fake). This study provides a text-based technique for identifying fake news. It goes through multiple steps, the first of which is assessing the properties of the incoming text and pre-processing it to clean it of all unimportant words. After that, work is applied to correct the word's spelling before adopting a proposed mechanism to summarize the texts by identifying the essential words and keeping all their repetitions in the resulting summary text to benefit from the most significant possible number of words when passing them to the word merging technology, using the FastText embedding methodology. This is followed by determining the most relevant textual qualities that separate legitimate news from fraudulent news using a suggested deep-learning method consisting of two convolutional layers connected by a self-attention mechanism layer. Before the outcome, the retrieved features are sent into the BiLSTM network, a type of recurrent neural network that is particularly effective in processing and classifying sequential data. The WelFake, ISOT, and TI-CNN datasets were used, and we achieved an accuracy of 90%, 97%, and 89%, respectively.