Abstract:

Abstract The sentiment analysis task is more complex considering the lack of relevant information in brief texts. Deep neural networks, like as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have been widely employed to extract information from data sentiment in recent years, with surprisingly good results. Though CNN can efficiently retrieve comparatively high features employing convolution and max-pooling layers, it cannot understand relationships' sequences. Parallelly bidirectional RNN models can extract contextual information and fail to extract local features. In this paper, integrated CNN and RNN models for sentiment analysis are examined to have the advantages of CNN's coarse grain local feature extraction and long-distance dependencies of RNNs. Particularly bidirectional LSTM and GRU networks associated with the convolution and max-pooling layer are used for sentiment analysis in SST-2 and movie review datasets. Two pre-trained word embedding techniques glove and word2vec are used. Experimental findings show that max performance is achieved at 93.44% for SST-2 and 95.42% for the movie review dataset using CNN BiGRU word2vec and CNN BiGRU glove, respectively.

Keywords: