Emotion recognition of social media users based on deep learning

Overall framework

This article proposes an improved LDA and CNN-BiLSTM emotion classification model. The overall framework is shown in Fig. 2 (Rhanoui et al., 2019; Liu et al., 2020).

The specific implementation process is as follows:

1. Through Python and XPath technology, the user-defined web crawler is used to collect the comment information of microblog social media public opinion events, including “Joy,” “anger,” and “sadness,” and store them in the local CSV file.

2. The data preprocessing of comment text includes Jieba Chinese word segmentation, stop word filtering, special character deletion, duplicate comment deletion, comment annotation, etc.

3. The deep features of the word model are extracted from the model and used as the input feature of the word model.

4. The CNN-BiLSTM model is constructed, and CNN is used to extract the key features of the text, and LSTM captures the long-distance semantic features. Finally, the Softmax classifier calculates the sentiment tendency of social media public opinion event comments to complete the emotion classification. The output results correspond to “joy,” “anger,” and “sadness,” respectively.

Emotional feature extraction

A feature extraction method is constructed using the LDA model and emotion dictionary. Figure 3 shows its specific implementation process.

1. The LDA model is used to extract the emotional feature words of different review texts, a topic model proposed by Blei & Jordan (2003).

2. Feature extraction is carried out by using an emotional vocabulary ontology database. At the same time, combined with the word frequency statistics and manual tagging of inspirational feature words, the emotional lexicon for microblog social media is constructed.

3. The feature extraction task is completed by the emotional lexicon. This operation can extract high-quality feature words with emotional color in different reviews and support the subsequent deep learning model to implement emotion classification tasks.

CNN model

A convolutional neural network (CNN) mainly comprises convolution and pooling layers. In this article, a three-layer CNN is constructed to extract the key features of comment text with sentiment events. The convolution layer will receive the emotion feature matrix of n × d, and the convolution process is shown in Eq. (1). (1)${\displaystyle h_i^d=f\left(w_d\times V_i+b_d\right)}$

In the formula, f represents the activation function, and the ReLu function is usually used to accelerate the training convergence speed. h^d represents the feature of comments on Weibo social media after vector convolution processing; w_d represents the convolution kernel of size d; V_i represents the word vector of the input layer. b_d represents the offset item. This convolution operation can effectively generate local feature sets, as shown in Eq. (2). (2)${\displaystyle H_d=\left\{h_1^d,h_2^d,\dots ,h_{n-d+1}^d\right\}.}$

The pooling layer can compress the size of text feature vectors and model parameters, maximizing emotional feature retention. Its calculation formula is shown in Eq. (3). (3)${\displaystyle s_i=max\left\{H_d\right\}.}$

Filters with convolution kernels of 2, 3 and 4 were constructed to extract key features of Weibo comment text, and then their output vectors were input into the BiLSTM model.

BiLSTM model

The bi-directional long short-term memory (BiLSTM) model is a variant of the recurrent neural network, which extracts features from the front and back directions to capture the long-distance dependency and context semantic features. This article extracts the emotional features of public opinion event reviews.

The network structure of the BiLSTM model is shown in Fig. 4,

Through the transmission of state to enhance the subject information and to effectively capture emotional characteristics such as “like” and “haha,” “uncomfortable,” and “pray,” the calculation formulas are shown in Eqs. (4) to (6). (4)${\displaystyle \overrightarrow{h_t}=f\left(w_1\times x_t+w_2\times \overrightarrow{h_{t-1}^{\to }}\right)}$ (5)${\displaystyle \stackrel{⃖}{h_t}=f\left(w_3\times x_t+w_5\times \overrightarrow{h_{t+1}}\right)}$ (6)${\displaystyle y_t=g\left(w_4\times \overrightarrow{h_t}+w_6\times \stackrel{⃖}{h_t}\right)}$

where, ${\overrightarrow{h}}_t$ represents the state of the forward LSTM layer at time t, and $\stackrel{⃖}{h_t}$ $\stackrel{⃖}{}$ represents the state of the backward LSTM layer at time t. x_t represents the input word vector; w₁ to w₆ represents weight parameters; f represents activation function; y_t is the final output of the bidirectional LSTM layer. Finally, the vector obtained by the BiLSTM model is input into the Softmax classifier to realize the emotion classification. That is, the emotion categories of “joy,” “anger,” and “sadness” are predicted.

Data acquisition

Building a web crawler using Python and XPath technologies to gather comment data on public opinion events on Weibo (https://weibo.com) is the experimental data for this research. 200,000 data sets with emotive color were created after data cleaning and preprocessing. Three data sets—one for each emotion (joy, anger, and sadness) were randomly split into training, test, and validation sets. The ratio of training set, test set and validation set was 3:1:1. Table 1 displays the distribution of the data.

Evaluation index

A confusion matrix is a standard tool used to evaluate unbalanced data, as shown in Table 1. It can obtain classification accuracy, precision, recall, and F1.

For sentiment analysis of Weibo social media comments, F1 and accuracy are used for experimental evaluation in this article, and the calculation process is shown in Eqs. (7)–(10).

(7)${\displaystyle \text{Precision}=\frac{TP}{TP+FP}}$ (8)${\displaystyle \text{Recall}=\frac{TP}{TP+FN}}$ (9)${\displaystyle F_1=\frac{2\times \text{Precision}\times \text{Recall}}{\text{Precision}+\text{Recall}}}$ (10)${\displaystyle \text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN}}$

Among them, precision is used to evaluate the percentage of emotion classification correctly predicted as the percentage of specified category in the number of anticipated category reviews. The recall is used to assess the percentage of emotion classification correctly predicted in the number of emotion reviews of the category. F₁ is a weighted harmonic mean of precision and recall.

Experimental process

This article used the LDA model and emotion dictionary. The n_topic of the LDA model was set as three, corresponding to “joy,” “anger,” and “sadness,” respectively. This operation allows unnecessary noise feature words to be effectively filtered, and feature words will be more emotional after processing, which provides good support for the subsequent emotion classification of the CNN-BiLSTM model. The specific process is as follows:

1. Chinese word segmentation and data cleaning (including stop word filtering and special character cleaning) extract feature words that only retain semantic value information.

2. The LDA model is used to extract the emotional feature words of “joy,” “anger,” and “sadness,” and the emotional feature words of other comments are added by combining the emotional vocabulary ontology database and manual annotation.

3. The emotion features are extracted from the CNN-BiLSTM model.

After feature extraction based on the LDA model and emotion dictionary, this article constructs the CNN-BiLSTM model and realizes the social media sentiment analysis experiment. The hyper-parameters of the model are shown in Table 2.

In addition, the Epoch of the model is 200. A Dropout layer is added to prevent overfitting. To avoid the influence of one abnormal experiment result, the whole experiment result is the average value of ten experiment results. At the same time, it is compared with classical machine learning models (including logistic regression, SVM, random forest, KNN, naive Bayes, AdaBoost) and deep learning models (including LSTM, BiLSTM, Gru, BiGRU, CNN, TextCNN).

Results of emotional feature word extraction

As shown in Fig. 5, the emotional feature word extraction model proposed in this study completes the social media sentiment analysis task. These words are social media lingo and frequently have particular meanings. The “anger” category includes terms like “without,” “problem,” “death,” “real,” “pathetic,” “serious,” and “angry,” as well as Internet terms like “TMD,” “HeHe,” and “vulnerable.” These feature words effectively reflect the public’s anger on public opinion events, and feature extraction based on LDA model and emotion dictionary can effectively enhance the result of eliciting emotion. The key emotional features of the “sadness” category include “pathetic,” “distress,” “silence,” “pray,” “blessing,” and “pity.”

Comparison of different models

Figure 6 shows that the F1 value of the proposed model is 0.89, and the accuracy rate is 0.87. The experimental results are better than the existing machine learning and deep learning models.

By comparing the changing trend of the F1 score between this method and other methods, it can be found that this method is 0.1878, 0.1939, 0.1902, 0.2671, 0.1419 and 0.2194 higher than logistic regression, SVM, random forest, KNN, Naive Bayes and AdaBoost, respectively. In addition, it is 0.0795, 0.0489, 0.0858, 0.0572, 0.0491 and 0.0394 higher than LSTM, BiLSTM, GRU, BiGRU, CNN and TextCNN, respectively. It can be seen that proposed method is 0.1747, 0.1789, 0.1937, 0.2597, 0.1477 and 0.2092 higher than logistic regression, SVM, random forest, KNN, Naive Bayes and AdaBoost, respectively. However, compared with LSTM, BiLSTM, GRU, Bi- GRU, CNN and TextCNN, the accuracy of the proposed model is improved by 0.0769, 0.0382, 0.0855, 0.0564, 0.0530 and 0.0459 respectively.

In addition, this article compares the F1 values of the fusion of the LDA model and emotion dictionary in different methods, and the experimental results are shown in Fig. 7.

The results show that the fusion of the LDA model and emotion dictionary indicates better performance in sentiment analysis of social media reviews. The F1 value of six machine learning models is increased by 3.66%, and that of seven deep learning models is increased by 1.84%, which shows that the effective extraction of emotional feature words can improve the effectiveness of the classification model to a certain extent. It can fully realize the sentiment analysis of the comments on social media public opinion events, better perceive the public sentiment and predict the emotional trend. It can give full play to the advantages of multilevel and multi-scale feature extraction networks in feature extraction, and can better extract word-level, phrase-level and sentence-level features to ensure the adequacy of feature extraction

Analysis of different emotional categories

Figure 8 shows the analysis results of varying emotion categories. It can be seen from the figure that the F1 value of “joy” is the highest, which is 0.9062, followed by “sadness” and “anger.” On the one hand, there are a large number of samples of the type of “joy,” and on the other hand, there is a phenomenon of partial integration of the emotional feature words of “anger” and “sadness.” However, the experimental results still effectively prove the method’s effectiveness in this article, which can conduct high-quality emotional trend analysis on the comment information of social media. Automatically distinguishes between different types of emotion such as “joy,” “anger,” and “sadness.”