*5.3. Performance Analysis*

Here, the impacts of different parameters on our model performance are analyzed.

• The length of the *maxlen*

In the initialization and padding of word vectors, the parameter, maxlen, was set up to determine the length of the words that are chosen to represent the article. As for SST-1 and SST-2 datasets, the average length of articles are 18, and the length is so short that it is difficult to obtain the different influences of the article length. Therefore, the THUCNews dataset is selected for an experiment to find out the effect of article length. Figure 4 shows that different lengths of the articles result in different performance.

The best result, 93.77%, occurs when 80 words are employed to represent the article, which is also deemed as the closest length to the average article length of the dataset. Once the maxlen is far greater than the average length of the articles, then the accuracy will decrease greatly because many more zero vectors will exist in the vectors of the article.

**Figure 4.** Accuracy vs. article length.

• The size of Convolutional Filter

In Figure 5, different convolution filter configurations are adopted to present the prediction accuracy on the question classification. As for the horizontal axis, the number indicates convolutional kernel size, and bar charts of five different colors on each filter size represent different dimensions of the convolution output. For example, "S8" means that the size of the kernel is 8 while "F128" denotes that the dimension of the convolution output is 128. It is quite obvious that the dimension of 64 outperforms the other dimensions and the best result, the accuracy of 93.55%, occurs when the window size is 7.

**Figure 5.** Accuracy vs. different filter configuration.

#### **6. Conclusions**

This paper mainly introduces a combined model called BLSTM-C that is made up of a bi-directional LSTM layer and a convolutional layer. The paper also shows its ability in information learning from both previous context and future context, as well as its competence to extract features. It is shown by the experiment results that this model performs remarkably on the tasks related to Chinese News classification, and it also outperforms CNN, RNN and other models on sentiment classification. Moreover, by running the experiments on similar dataset in Chinese and English, it is found out that our BLSTM-C model may have better performance in the Chinese language because the improvement shown in the Chinese language is more significant. To obtain better performance in Chinese news classification, the suitable parameters for this model are also explored and it is found that it is helpful to improve the results by setting the maxlen closest to the average article length and adopting a suitable window to detect the features.

Furthermore, the following suggestions may lead to better performance in future work. Firstly, it will be an interesting idea to deepen the neural network layer. Research on text classification by employing 29 convolutional layers gets satisfying results. Moreover, it is also common for Long Short-term Memory (LSTM) to be multi-layer. Secondly, the article will be more reasonable if every location, number, name and some other meaningless words are replaced with placeholders like '[location]', '[number]','[name]'. This operation will enable the article to be clearer for computer systems.

Therefore, in future work, these modifications will be tried in our experiment to see if it will achieve better performance.

**Funding:** This research was funded by the Natural Science Foundation of Shanghai Grant No. 16ZR1401100, and the National Key R&D Program of China, Grant No. 2017YFB0309800.

**Conflicts of Interest:** The authors declare no conflicts of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

MDPI St. Alban-Anlage 66 4052 Basel Switzerland Tel. +41 61 683 77 34 Fax +41 61 302 89 18 www.mdpi.com

*Future Internet* Editorial Office E-mail: futureinternet@mdpi.com www.mdpi.com/journal/futureinternet

MDPI St. Alban-Anlage 66 4052 Basel Switzerland

Tel: +41 61 683 77 34 Fax: +41 61 302 89 18