The XLNet-CNN-BiLSTM-Att model consists of four main components, namely the XLNet layer, CNN layer, BiLSTM layer, and Attention layer. The model architecture is shown in Figure 1. The overall process of the model is as follows. First, the XLNet layer is used to obtain word vectors containing contextual semantic information. Second, then local semantic features are extracted using CNN. Then, BiLSTM extracts contextually relevant features. Finally, the attention mechanism is introduced to assign weights to the extracted features to highlight the key information. The model utilizes a softmax classifier for sentiment classification.

XLNet layer: The XLNet layer is used to vectorize text. The text sequence X ( X 1 , X 2 , … , X n ) with input length n is preprocessed, and the word vector corresponding to each word g ( g 1 ( 2 ) , g 2 ( 2 ) , … , g n ( 2 ) ) is learned through the XLNet model. In which the word vector g encodes its context information and its own position information. The word vectors learned by the XLNet layer will be used as input data for the CNN layer.

CNN layer: The purpose of the CNN layer is to learn the local features of a text. It inputs the word vector calculated by the XLNet layer into the convolution layer and completes the feature extraction of the input text through three filters. The local feature matrix C_j obtained after the convolution layer is input into the pooling layer, and down-sampling is performed by the MaxPooling technique to obtain the optimal solution of the local value.

BiLSTM layer: The BiLSTM layer aims to learn the contextual information features of words in sequence data. The feature vector calculated by the CNN layer is input into the hidden layer of BiLSTM. A standard LSTM computes a sequence of hidden layer vectors from one direction, while a bidirectional LSTM computes from two different directions and finally combines the results from both directions for output. The forward hidden layer vector is h → , the backward hidden layer vector is h ← , and the output vector of the hidden layer at the moment is v i = h ← +   h → .

Attention layer: The Attention layer is to assign different weights to the feature vectors. The Attention mechanism assigns different weights to the feature vectors selected by the BiLSTM layer according to different degrees of influence a (a₁, a₂, …, a₃), so as to improve the attention of the model to important features. Finally, the text sentiment is calculated by the softmax activation function.

The common metrics used to evaluate the performance of sentiment analysis models are Precision (P), Recall (R), F1-score (F₁), Accuracy (Acc) and Loss rate (L).

A c c = T P + T N T P + T N + F P + F N × 100 % P = T P T P + F P × 100 % (1)

R = T P T P + F N × 100 % F 1 = 2 × P × R P × R × 100 % L = − [ log y ^ + ( 1 − y ) log ( 1 − y ^ ) ] (2)

In this experiment, three models were chosen as experimental control models: Word2Vec-BiLSTM, XLNet and XLNet-CNN. The sentiment analysis experiments were conducted on three test events, including Henan event, Tianjin event, and Anhui event. Figure 2 depicts the dynamic changes in the accuracy and loss rate of the model training over the 10 epochs. According to the accuracy and loss rates of the validation set data during the training phase, we may conclude that the model proposed in this paper outperforms the other five models.

In this study, the optimal parameter values of XLNet-CNN-BiLSTM were utilized to classify Internet users’ sentiments, and the results are presented in Table 3. The average accuracy of the three test events is 94.12%, with the Anhui event having the highest accuracy at 94.75%. This demonstrates that the XLNet-CNN-BiLSTM model proposed in this research is capable of achieving outstanding performance in Internet public opinion sentiment analysis. Besides, the model outperformed the other five models in terms of accuracy in all three events tested. This implies that the dynamic text representation approach is superior to the static text representation method Word2Vec for classification tasks.

In this paper, the emotional value of online public opinion text on food safety is constructed in the same way as the bullish index used by Antweiler in the stock market, which is used to reflect the emotional trend of netizens in public opinion events over a period of time, and can integrate emotional information. The calculation method of the bullish index combines the bullish and bearish sentiment of the investors in the stock market. It is a classic calculation method in the

measurement of investor sentiment in the stock market. It is similar to the positive and negative emotional tendencies of the Internet users of the food safety network public opinion on public opinion events. Therefore, this method is used for reference to construct the emotional value of the food safety network public opinion in this paper. Referring to the bullish index formula, the emotional value calculation formula is S t = ln [ ( 1 + n t p o s ) / ( 1 + n t n e g ) ] , n t p o s indicates the number of blog posts with positive emotions in the food safety network public opinion event in a certain period, n t n e g indicates the number of posts with negative emotions in a certain period.

In order to use the time series model to predict the emotional trend more accurately, the emotional value calculated for each blog article is stored in chronological order. The emotional value data of “Henan Event”, “Anhui Event” and “Tianjin Event” are selected as the data source of the emotional trend prediction model, Take the average of the emotional values of the online public opinion blog posts corresponding to these food safety events every 30 minutes and give the emotional value sequence { S 1 , S 2 , … , S t } that constitutes the corresponding time point. The emotional value time series is divided into training sets and test sets. A set of training samples is used to train LSTM to generate memory, and then the prediction set is input into the training LSTM to obtain the output of the prediction set.

Because the emotional tendency of blog posts is easily affected by many factors such as opinion leaders, Internet users’ relevance and the nature of events, it usually contains a lot of noise, so the constructed temporal emotional value also contains noise. The wavelet analysis can refine the signal analysis in many degrees through the operation functions such as expansion and translation. It can eliminate the noise in the time series data, but also cannot change the characteristics of the original time series data. Therefore, this paper uses wavelet denoising to eliminate the noise components in the time series data of the emotional value of the public opinion of the food safety network, reduce the interference impact of noise on the LSTM model, and improve the prediction function of the model. Wavelet analysis can decompose each layer of input data into a low frequency signal and a high frequency signal, and can further decompose the low frequency signal into a low frequency signal and a high frequency signal, and so on.

The LSTM model can mine the information in the long time series data. Compared with the problem of gradient disappearance or explosion that may occur in RNN, LSTM has effectively solved the problem of long-term dependence and gradient disappearance by deliberately improving the internal structure of RNN and adding three gating units, namely input gate, forgetting gate and output gate. The purpose of this paper is to predict the emotional trend of food safety network public opinion through a neural network model. Considering the impact of historical emotional value time series data on the emotional trend, LSTM can effectively use historical information and has advantages in processing long-term dependent data, and build an emotional value time series prediction model based on LSTM model.