**5. Results**

As shown in Table 1, our model is compared with 14 well-performed models from different tasks. One of the tasks is sentiment classification (SST-1, SST-2) while the other one is category classification (THUCNews).

## *5.1. Overall Performance*

Both SST-1 and SST-2 datasets are employed to compare the performance of different methods. As shown in Table 1, our model is compared with some well-performed models from different areas, such as Support Vector Machine(SVM), Recursive Neural Network, Convolutional Neural Network,

and Recurrent Neural Network. Specifically, for the Recursive Neural Network, what are chosen include MV-RNN: Semantic Compositionality through Recursive Matrix-Vector Spaces [18], RNTN: Recursive Deep Models for semantic compositionality over a sentiment treebank [2], and DRNN: Deep Recursive Neural Networks for compositionality in language [19]. For CNN, what are chosen include DCNN: a CNN for modeling sentences [6], CNN-nonstatic and CNN multichannel: Convolutional Neural Networks for sentence classification [10], and Molding-CNN: Molding CNNs for text, including nonlinear and non-consecutive convolutions [20]. For Recurrent Neural Networks, what are chosen include RCNN: Recurrent Convolutional Neural Networks for Text Classification [20], S-LSTM: Long Short-term Memory over recursive structures [21], BLSTM and Tree-LSTM: improved semantic representations from tree-structured Long Short-term Memory networks [12]. For other baseline methods, we use a Support Vector Machine, n-gram bag of words and a Paragraph Vector. In addition, LSTM and B-LSTM were also implemented by us for further comparison of category classification on our Chinese news dataset.


**Table 1.** Comparison with baseline models on Stanford Sentiment Treebank and THUCNews.

The table above shows that our BLSTM-C model achieves remarkable performance in two out of three tasks (numbers in bold represent the best results). In sentiment classification, our BLSTM-C model gets the best result in the SST-2 dataset while molding-CNN achieves the best performance in the SST-1 dataset. Although our model fails to beat the state-of-art ones, it still obtains an acceptable result which means that the model is feasible for various scenarios.

As for the classification of text category, our model outperforms other well-performed models, achieving outstanding results. Through the comparison between our model and single-layer LSTM, B-LSTM and LSTM models, it is found that our model does combine the advantages of both LSTM and CNN. Apart from successfully learning long-term dependencies, it extracts features from text, leading to better results. Although almost no human-designed features are employed in our model, it beats the state-of-the-art SVM that highly requires engineered features.
