A bigura-based real time sentiment analysis of new media

Construction of the attention convolutional neural network conditional random field word segmentation model

The attention convolutional neural network conditional random field word segmentation model is a deep learning model that combines the attention mechanism and convolutional neural network to improve accuracy and efficiency. The attention mechanism can focus on key vocabulary with important information. Convolutional neural networks can effectively capture the contextual relationships between words. Through this combination, the model can more accurately segment new media texts, providing more accurate input information for subsequent sentiment analysis and public opinion monitoring (Li & Xu, 2020; Seokhoon, Jihea & Young-Sup, 2023). The overall framework diagram of the attention convolutional neural network conditional (ACNNC) model is displayed in Fig. 1.

In Fig. 1, the ACNNC model is composed of five levels. The order is an embedding layer, attention layer, CNN layer, fusion layer, and CRF layer. Firstly, the embedding layer generates 128-dimensional word vectors through training, which will serve as inputs for subsequent levels. Then, the word vector is input to the attention layer and CNN layer. The former is responsible for learning the overall features of the sequence. The latter learns the positional and local features of the sequence. Next, the fusion layer integrates the overall, local, and positional features obtained from the above learning. Finally, the integrated features are decoded at the CRF layer to complete the construction of the model (Hou, Wang & Wang, 2023).

The embedding layer converts the ith word in a sentence into a word vector. Then, it is trained to obtain the word vector matrix. The size of the word vector matrix is determined by the effective length of the vocabulary in the corpus dataset and the dimension of the input word vector. Each word can be transformed into a corresponding word vector representation through a word vector matrix. The embedding layer is displayed in Eq. (1). (1)${\displaystyle e_i=X\times v_i}$

In Eq. (1), X is the word vector matrix. $\left|N\right|$ is the effective length of the vocabulary in the corpus dataset. v_i is a One-hot vector of size $\left|N\right|$. The sentence x isconverted to S, as shown in Eq. (2). (2)${\displaystyle S=\left(e_1,e_2,\dots ,e_n\right)\in R^{\left|N\right|\times d}.}$

In Eq. (2), d stands for the dimension of the input word vector. The individual attention is shown in Eq. (3). (3)${\displaystyle hea{d}_i=Attention\left(Q{W}_i^Q,K{W}_i^K,V{W}_i^V\right).}$

In Eq. (3), Q, K, V stand for the same value. The attention layer adopts a multi-head self-attention mechanism to reduce the possible random errors caused by the single-head attention mechanism and further improve the accuracy. Unlike conventional attention mechanisms, the dependency of self-attention mechanisms lies in themselves. That is, the query, key, and value are all the same value. The multi-head self-attention mechanism is shown in Fig. 2.

In Fig. 2, the query, key, and value are first linearly mapped, followed by eight individual attention calculations. Finally, the results are concatenated. The characteristic of this self-attention mechanism is that the queries, keys, and values are all the same value. In this way, the multi-head self-attention mechanism can improve model accuracy. In the CNN layer of the model, only the convolutional layer and output layer are included. Input sentences processed by the embedding layer are fed into the convolutional layer for convolutional operations (Fang et al., 2022; Yang & Song, 2022). The convolution operation is shown in Eq. (4). (4)${\displaystyle x_i^k=\sigma \left(\sum \left(X\left[i:i+k\right]\circ H_k\right)+b\right).}$

In Eq. (4), σ stands for the activation function. ∘ is a point multiplication operation. $X\left[i:i+k\right]$ is a sequence of word vectors. H_k is a convolutional kernel of size k. The fusion layer integrates the features of the attention layer and the CNN layer, generating feature parameters containing overall, local, and positional information and then inputting them into the CRF layer. This model utilizes a vector concatenation strategy for feature fusion. CRF is a discriminative undirected graph model that can consider changes in data content and labels. It is widely used in tasks such as Chinese word segmentation, named entity recognition, and part of speech tagging. The processed sentences in the model are output through the fusion layer and then input into the CRF layer to obtain the results of sequence annotation. After indexing and standardizing the scores of sequence annotations, the final probability value is obtained. The calculation of S_t is shown in Eq. (5). (5)${\displaystyle S\left(y\left|x\right.\right)=\sum _{i,k_1}{\theta }_{k_1}{t}_{k_1}\left(y_{i-1},y_i,x,i\right)+\sum _{i,k_2}{\mu }_{k_2}{s}_{k_2}\left(y_i,x,i\right).}$

In Eq. (5), O_t represents the character after passing through the fusion layer. $t_{k_1}\left(y_{i-1},y_i,x,i\right)$ is the transfer feature function. $s_{k_2}\left(y_i,x,i\right)$ is the state feature function. θ_k1 and μ_k1 are the model parameters to be estimated, respectively. The probability value is shown in Eq. (6). (6)${\displaystyle P\left(\hat{y}\left|x\right.\right)=\frac{1}{Z\left(x\right)}\left({\theta }_{k_1}{t}_{k_1}\left(y_{i-1},y_i,x,i\right)+\sum _{i,k_2}{\mu }_{k_2}{s}_{k_2}\left(y_i,x,i\right)\right).}$

In Eq. (6), $P\left(\hat{y}\left|x\right.\right)$ is the probability value. The normalization factor is the sum of all possible marker sequences, as shown in Eq. (7). (7)${\displaystyle Z\left(x\right)=\sum _{y}exp\left(\sum _{i,k_1}{\theta }_{k_1}{t}_{k_1}+\sum _{i,k_2}{\mu }_{k_2}{s}_{k_2}\left(y_i,x,i\right)\right).}$

Equation (7) Z is the normalization factor. The task of word segmentation sequence annotation is to select the most likely marking sequence for each character in a given sentence. The markers include the Begin (B), Middle (M), End (E), and Single (S). The sequence annotation model is used for training and prediction, calculating conditional probabilities. Then, the sequence with the highest probability is output to obtain the segmentation result. The process diagram of the word segmentation sequence is shown in Fig. 3.

In Fig. 3, the word segmentation process includes four stages: preprocessing, model training, prediction, and post-processing. During the preprocessing stage, special symbols, punctuation, and stop words are removed to reduce model complexity. During the model training phase, a large amount of annotation data is used to train sequence annotation models, such as HMM, BiLSTM, etc., to generate the most likely label sequence for each character or sub-word. In the prediction stage, a trained model is used to segment new sentences and output a tag sequence representing the most likely markers for each character or sub-word. The post-processing stage converts the tag sequence into the actual segmentation result. The specific methods for each stage depend on task requirements and resource constraints.

Construction of an emotional analysis model based on aspect information

On the basis of the ACNNC model, this study further explores the construction of an emotion analysis model based on aspect information. Emotional analysis faces complex contextual and subjective challenges in extracting emotional tendencies from texts. To obtain emotional information more accurately, aspect information is introduced for more in-depth text fine-grained emotional analysis. The BiGRUA model flowchart is shown in Fig. 4.

In Fig. 4, the construction of the model includes five stages, preprocessing, input layer, Bidirectional Gated Recurrent Unit (BiGRU) layer, attention layer, and classification layer. Firstly, in the preprocessing stage, the original text data is cleaned and standardized to remove irrelevant information and extract effective features. Next, the input layer trains the data through the Word2vec model to obtain a word vector matrix. It can transform the original text data into a sequence of word vectors with certain semantic information. Next, the training set and corresponding training labels are fed into the BiGRU layer. At this stage, the model learns contextual information in the text through forward and backward GRU and updates model parameters to better capture semantic information in the text. The output of the BiGRU layer is fed into the attention layer. The attention mechanism focuses on the most important parts of the text, thereby extracting more representative features. Finally, the task of the classification layer is to classify emotions based on the output of the attention layer. Through this process, the model can extract useful features from the original text data and perform accurate emotional analysis. The BiGRUA model framework diagram is shown in Fig. 5 and its clever setting is compared with other models.

In Fig. 5, BiGRUA is proposed for sentiment analysis of text containing multiple words, which integrates aspect information. It uses GRU instead of LSTM for feature learning. The attention mechanism is used to allocate aspect information weights. Finally, sentiment analysis is achieved through a softmax classifier. The BiGRU layer solves the gradient vanishing and long-distance semantic capture in recurrent neural networks by using GRU units in update and reset gates. It combines positive and negative implicit states, which can fully learn reverse semantics and obtain more complete contextual information. The BiGRU layer structure is shown in Fig. 6.

In Fig. 6, the unit structure of GRU plays a key role in the sentiment analysis model based on aspect information. Based on reset and update gates, the model can handle long-distance semantic dependencies and adapt to complex sentiment analysis tasks. The reset gate determines the old information that the model should retain or discard before further processing new information. The update gate is responsible for how to integrate new and old information to generate the latest status. This mechanism enables GRU to effectively alleviate the gradient disappearance while capturing long-term dependencies. In the emotion analysis model based on aspect information, the unit structure of GRU has crucial impacts on the accuracy of the model due to the special design and critical role. Based on the clever setting of reset and update gates, this model can handle the complex task of finding semantic associations in long-distance contexts, thus adapting to the complexity of sentiment analysis. This clever setting may be described from its connectivity with self-attention and Bayes condition probabilities. For a clear understanding, an the architecture of BiGRU is used for text emotion detection and visual emotion detection is given in the Fig. 5. The purpose of resetting the gate is to determine which old internal states should be retained or discarded before further processing new information. This is a key mechanism. GRU can flexibly adjust the internal state based on new input and context, thereby effectively capturing and understanding long-distance dependencies in text. The update gate is responsible for fusing new and old information to generate the latest status. This mechanism is implemented through a carefully designed gating mechanism. It can fuse and update new and old information appropriately based on the characteristics and context of the input text. This design enables GRU to effectively alleviate the gradient vanishing while capturing long-term dependencies, thereby improving the stability and accuracy of the model.

The output of the BiGRU layer at a time t is displayed in Eq. (8). (8)${\displaystyle h_t^{{}^{\prime }}=\left[{\overrightarrow{h}}_t,{\overleftarrow{h}}_t\right].}$

In Eq. (8), $h_t^{{}^{\prime }}$ is the output layer. The attention layer is to distinguish the importance of different parts in a sentence, as the influence of target words and viewpoint words on the emotional orientation in the text is different. The attention mechanism is applied to assign different weights to different parts, highlighting the contribution of different aspect words to emotional tendencies. The input is the output of the BiGRU layer, which is transformed into a new hidden vector through a multi-layer perceptron. Then it is calculated with the context vector to obtain the weight value. This context vector is a high-dimensional vector used to determine the importance of words in a sentence. The attention layer is displayed in Eq. (9). (9)${\displaystyle \begin{array}{c}{u}_t^{{}^{\prime }}=tanh\left(W_w{h}_t^{{}^{\prime }}+b_w\right)\hfill \\ {\alpha }_t=\frac{exp\left(u_t^{{}^{\prime }T}{u}_v^{{}^{\prime }}\right)}{\sum _{t}exp\left(u_t^{{}^{\prime }T}{u}_v^{{}^{\prime }}\right)}\hfill \end{array}.}$

Equation (9), $u_t^{{}^{\prime }}$ is the new hidden vector. α_t is the weight value. $u_v^{{}^{\prime }}$ is a high-dimensional vector used to determine the importance of words in a sentence. The average target word vector is shown in Eq. (10). (10)${\displaystyle V_{at}=\frac{1}{m}\sum _{i=1}^m{e}_i.}$

Equation (10) m is the target word in the sentence. V_atG is the target word vector. The fusion vector is shown in Eq. (11). (11)${\displaystyle P_i=h_t^{{}^{\prime }}+V_{at}.}$

In Eq. (11), $h_t^{{}^{\prime }}$ is the hidden vector of the output. P_i is the fusion vector. The emotional features of attention weights that integrate aspect information are shown in Eq. (12). (12)${\displaystyle C^{{}^{\prime }}=\sum _t{\alpha }_t{h}_t^{{}^{\prime }}.}$

In Eq. (12), C′ integrates the emotional features of aspect information attention weights. The classification layer inputs the features captured by the attention layer into the softmax classifier for classification. The activation function adopts the sigmod function, as shown in Eq. (13). (13)${\displaystyle \tilde{y}=softmax\left(W_{C^{\prime }}{C}^{\prime }+b_{C^{\prime }}\right).}$

In Eq. (13), W_C′ is the weight matrix. b_C′ is the offset value. $\tilde{y}$ is the classification result. The cross entropy loss function is shown in Eq. (14). (14)${\displaystyle L=-\frac{1}{N}\sum _{r=1}^m\sum _{q=1}^C{y}_{rq}log\left({\tilde{y}}_{rq}\right)+\lambda {∥\theta ∥}_2.}$

In Eq. (14), y and $\tilde{y}$ stands for the actual label values and predicted label values. C is the number of label categories. λ is the L₂ regularization coefficient.

New media sentiment analysis and public opinion monitoring technology analysis based on deep learning

The rapid development and widespread application of new media have enabled the rapid acquisition and dissemination of large-scale public sentiment tendencies and public opinion dynamic information. Deep learning, as a complex and powerful machine learning technology, has been widely applied in sentiment analysis and public opinion monitoring. The application effect of deep learning in new media sentiment analysis is discussed in detail. The performance of new media public opinion monitoring technology is compared, providing theoretical reference and practical guidance for research in related fields.

The application effect of deep learning in new media sentiment analysis

In the new media sentiment analysis, the application effect of deep learning has attracted much attention. By constructing precise models and algorithms, deep learning can accurately identify and understand public emotions, providing valuable emotional feedback for enterprises and organizations. Table 1 displays the parameters.

To better demonstrate the practical application effect of deep learning in emotional analysis, a study collected new media content from a certain self-media platform as the parent dataset and divided the dataset into two sub-datasets: text content and video content. Based on these two sub-datasets, experiments were conducted on binary and ternary classification, as shown in Fig. 7.

During the experiment, the study conducted multiple iterations based on different parameter settings and recorded the accuracy of the model after each iteration. All parameter settings and corresponding accuracy are shown in Table 1. In Fig. 7, in the binary classification experiment, the model achieved the highest recognition accuracy of 92.7% in text-content emotional expression and 86.9% in video-content emotional expression evaluation. The accuracy of recognizing emotions in text content gradually increased from 82.9% in the first iteration to the highest in the 11th iteration and fluctuated after that. The accuracy of video content emotion recognition fluctuates and increases, reaching its highest after 15 iterations. When analyzing new media content datasets, the highest accuracy can reach 91.4%. In the three classification experiments, the results were similar to those of the two classifications. The highest accuracy of text content emotion expression recognition is 90.8%, and the highest accuracy of the parent dataset is 89.0%. The accuracy of emotional expression evaluation in video content is relatively low. The highest accuracy of the ternary classification is lower than that of the binary classification. The reason may be that the increase in classification types led to a decrease in accuracy. Deep learning achieved high accuracy in emotional analysis of text content and video content, whether it is for binary or ternary classification. These experimental results fully demonstrated the superior performance of deep learning in new media sentiment analysis. It further confirmed the effectiveness of the parameter settings in Table 1. The accuracy of deep learning in new media sentiment analysis is shown in Fig. 8.

In Fig. 8, deep learning demonstrates significant advantages in the analysis of new media emotions. The classification accuracy is improved by 3.88% and 4.33% compared to Bayesian and KNN classifiers, respectively. After optimization, the accuracy is further improved to 6.74% and 6.97%. However, the performance of Bayesian models deteriorates when labelled data accounts for 40%. KNN performs excellently when labelled data accounts for 20%. Deep learning has improved the accuracy of sentiment classification and assisted in real-time public opinion monitoring, which is of great value for policy formulation and marketing.

Experimental results and analysis of deep learning in public opinion monitoring

The sentiments expressed via the Sina Weibo API center around the notion of Israel urging millions of individuals from northern Gaza to relocate southward. Following data preprocessing, the information is structured in JSON format. Using this scenario as a case in point, an examination is conducted on the comprehensive metrics encompassing topic popularity, sentiment ratio within comments, and the intensity of emotions expressed. In light of Israel’s relocation request, there has been a notable escalation in the intensity of public discourse, underscoring the necessity for vigilant public opinion monitoring, particularly in gauging both the fervor surrounding an incident and the prevailing emotional inclinations among the populace. The intensity of discussion, as reflected in the heat score pertaining to the topic “Israel requires millions of people from northern Gaza to migrate to the south,” is illustrated in Fig. 9.

Figure 9 depicts the emotional landscape surrounding the topic of “Israel requiring millions of people from northern Gaza to migrate to the south.” Initially, following Israel’s relocation proposal, there was a surge in public sentiment and the intensity of discussion. However, over time, the frequency of public discourse decreased, and emotions gradually subsided, returning to a state of calm. This underscores the significance of robust public opinion monitoring models, which necessitate keen observation of both the momentum of events and the prevailing sentiments among the public. The emotional orientation map pertaining to the topic “Israel demands millions of people from northern Gaza to migrate to the south” is illustrated in Fig. 10.

Figure 10 presents the emotional orientation map concerning “Israel’s demand for millions of people from northern Gaza to migrate to the south.” Initially, as the incident unfolded, the public expressed their perspectives with an expectation of Israel’s handling being reasonable, hence marking a low point in emotional inclination around the fifth timestamp. However, with Israel’s relocation request, public sentiment surged dramatically. Yet, by the 10th timestamp, due to various complex factors, emotions experienced a rebound. The evolution of the comprehensive score reflecting the intensity of emotional responses to the topic is depicted in Fig. 11.

In Fig. 11, initially, there is a noticeable level of heat, but no comments are recorded, resulting in a score of 0. As time progresses, attention towards the events escalates, leading to a surge in comments and an overall increase in the score reflecting emotional intensity. By the third timestamp, when Israel makes the relocation request, the score reaches its peak value, indicating heightened emotional engagement. However, as the event unfolds further, the score gradually diminishes, suggesting a gradual decrease in emotional intensity over time.