Suspicious classifying on opining in text is challenging task in Natural Language Processing (NLP). In NLP, standard approaches are ineffective at extracting multi-level properties of text sequences. However, relying solely on deep learning (DL) approaches, such as Deep Belief Networks (DBNs), Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and Two State Gated Recurrent Neural Networks (TS-GRU) ignore contextual meaning of the word. This study proposed TS-GRU-CBOW, which uses extraction to express text sentiments. The performance of the TS-GRU-CBOW model has been tested on various Twitter datasets and compared on different sentiment analysis techniques by employing GRU, LSTM, Bidirectional LSTM (Bi-LSTM), Bidirectional GRU (Bi-GRU), CNN, and TS-GRU. The proposed TS-GRU-CBOW is a feature attention-based method that uses sequential molding and word-feature seizing to identify and context sentiment analysis, as well as reducing data dimension using the attention mechanism. In this study, five sentiment datasets were used for sentiment analysis and were implemented with TS-GRU-CBOW model, whereby, reveals that TS-GRU-CBOW model outperforms other than baseline methods by accomplishing the highest accuracy of 85%, precision of 83.9, recall of 79.9, and F1-measure 86% score on yelp dataset.