The unavailability of annotated dataset for a low-resource Ewe language makes it difficult to develop an automated system to appropriately evaluate public opinion on events, news, policies, and regulations in the language. In this study, we collected and preprocessed a low-resourced document-level Ewe sentiment dataset based on social media comments on five (5) different topics. Additionally, we generate three (3) word embedding models (Global vectors, word-to-vector and continuous bag-of-words ) for exploiting sentiment representation based on the dataset. We further proposed a novel multi-channel two-dimensional (2D) convolutional neural network fuse with attention-based-bidirectional long-short term memory (MC2D-CNN+BiLSTM-Attn) to detect the exact sentiment feature from the Ewe document. The proposed method can efficiently describe the following emotions: anger, annoyance, happiness, surprise, and, sadness from the newly developed dataset. Extensive experiments indicate that MC2D-CNN+BiLSTM-Attn method marginally outperformed other known state-of-the-art methods. Results show that in detecting the precise sentiments from raw Ewe textual context, the BiLSTM incorporating Glove outperforms word2vec and CBOW embedding with an accuracy of 0.72714. Furthermore, Attn+BiLSTM and Multi-channel CNN methods incorporating word2vec embedding layer perform better than Glove and CBOW embedding with an accuracy of 0.8483 and 0.8965 while our proposed technique with the same word2vec embedding recorded 0.9493.