added to the model. The experimental data show that adding comment sentimental
features can improve the rumor detection effect of the model.
4.3.3 Experimental Comparison of Sentimental CNN-LSTM with Other
Models
This method is compared to traditional machine learning methods such as Naive Bayes,
SVM classifier model, and common deep learning model. The word vector model is
Glove, the word vector dimension is 250, and the rumor detection threshold is 0.3. The
experimental results are shown in Table 4.
From Table 4, we can see that the sentimental CNN-LSTM constructed in this
paper has the best effect in the comparative experiments of the above seven models.
The method based on deep learning is obviously better than traditional machine
learning method, and the performance of CNN-LSTM in deep learning method is
slightly better than other models. The validity of the sentiment CNN-LSTM proposed
in this paper is proved by comparative experiments.
Fig. 4. Change of accuracy in CNN-LSTM training
86.00%
88.00%
90.00%
92.00%
94.00%
96.00%
senƟmental CNN-LSTM
CNN-LSTM
Accuracy
Precision
Recall
F1
Fig. 5. Influence of sentimental features on model performance
154
S. Lv et al.