Along with the appearance and development of many social networking platforms in many different fields, there is an explosion of freely shared information. The benefits of these sources of information depend on whether the information is real or fake. The problem today is that we cannot distinguish factual information from fiction when receiving information on social networking platforms. To solve this problem, many different approaches to detect fake news have been proposed with promising performance, most notably approaches based on shared textual content. However, previous methods based on text content focus too on extracting features within the news’s leading content and consider the news title only as a relevant feature. This is not true for the habits of some users, who only read the news title without paying attention to the content. To overcome this limitation, we propose a dual BiLSTM-based multimodal method to detect fake news. The proposed method includes three main elements: (i) a BiLSTM-based fake headline extractor to capture and represent features of context and semantics on the headlines, (ii) a BiLSTM-based fake body extractor to capture and represent features of context and semantics on the news bodies, and (iii) a dual BiLSTM-based multimodal fake news detector to combine two fake headline and body features based on early fusion mechanism. The difference in the proposed method is that we do not consider the headline simply as one of the features; we view it as an independent text containing important features of context and semantics. The proposed model was tested on ISOT and FA-KES datasets and achieved better results than previous methods regarding the same content-based multimodal approach.