The microblog sentiment analysis based on latent dirichlet allocation and deep learning approaches

Abstract

To study the application of convolutional neural networks (CNN) in microblog sentiment analysis, a microblog sentiment dictionary is established first. Then, latent Dirichlet allocation (LDA) is proposed for user forwarding sentiment analysis. The sentiment analysis models of CNN and long short-term memory network (LSTM) are established. Experiments are conducted to verify the application effect. The main contributions of this work encompass the establishment of a sentiment lexicon for Weibo, the optimization of two sentiment analysis models, namely CNN and LSTM, as well as the comparison and analysis of the performance of three sentiment analysis approaches: CNN, LSTM, and LDA. The research findings indicate that the CNN model achieves a prediction accuracy of 78.6% and an actual output precision of 79.3%, while the LSTM model attains a prediction accuracy of 83.9% and an actual output precision of 84.9%. The three analysis models all have high sentiment analysis accuracy. Among them, LDA analysis model has the advantages of universality and irreplaceable in text classification, while LSTM analysis model has relatively higher accuracy in sentiment analysis of users forwarding microblog. In short, each sentiment analysis model has its own strengths, and reasonable allocation and use can better classify microblog sentiment.

Keywords

Convolutional neural networks (CNN) model deep learning human-computer interaction latent dirichlet allocation (LDA) model long short-term memory network (LSTM) model microblog sentiment

1. Introduction

Microblog is a short form of expression of life emotion in blog which has become popular in recent years. Short real-time information is shared at any time through microblog, spreading widely and forming a social network platform [1]. It can not only release various messages quickly and at any time using microblog, but also carry out real-time interaction among users in the member network. The reason why microblog is favored by users is that the requirements of microblog users is relatively low compared with traditional blogs, which makes microblog show an explosive development trend. It has developed into one of the most popular social networking places on the Internet with a very fast trend [2, 3]. In view of the fast speed of information transmission, strengthening communication with customers through technical research and development and market operation on Weibo has gradually become a new choice for enterprise development. Microblog will create a novel media atmosphere for users, including sound, text, images, mobile clients, and other hyperspace environments [4].

Microblog users can release microblog media in the form of SMS at anytime and anywhere, allowing all microblog visitors or their own designated blogs to browse the microblog content. Therefore, the introduction of microblog into human-computer interaction education has a great impact on the discourse power [5]. According to statistics, in 2016, more than 1,100 universities opened official Weibo accounts, and as of June 2018, the number of registered users of Sina microblog has exceeded 330 million [6, 7]. There is no doubt that microblog plays a positive role in our life, work, and education. However, with the growing maturity of microblog “our media”, all kinds of network emergencies will continue to spread and ferment on microblog and other “our media”, generating massive rumors and false information. In addition to information asymmetry and public perception deviation, once public opinion breaks out, there will be massive public opinions affecting social stability and unity. There will be a serious impact on the formulation of public policies in China, and the risk management of various departments will be extremely challenging [8]. Therefore, timely analysis of users’ emotions when forwarding microblogs, which aims to curb the widespread spread of negative public opinions, is of great significance to the social stability of China.

Latent Dirichlet allocation (LDA) model and deep learning model based on convolutional neural networks (CNN) model and long short-term memory network (LSTM) model are proposed innovatively. The microblog sentiment dictionary is established for the three models, the experimental environment of the deep learning model is analyzed. Then, the initial microblog text is obtained through web crawler. The main logical framework of this work is as follows. In Section 1 of the study, an introduction is provided on the background of sentiment analysis on Weibo and research methods using deep learning. In Section 2 of the study, an analysis is conducted on the current research status of deep learning and the utilization of neural networks in sentiment analysis. Section 3 of the study organizes the sentiment analysis using Dirichlet sentiment analysis and deep learning models. In Section 4 of the study, experimental design and analysis of experimental results are carried out through the configuration of experimental datasets and environments. Section 5 of the study derives research conclusions through induction and summarization. The research has practical reference value for promoting the development of convolutional neural network models in the field of sentiment analysis.

2. Literature review

2.1 Deep learning status

Human-computer interaction and deep learning technologies have been applied in many fields with the rapid development of artificial intelligence technology in the Internet era [9]. Interactive teaching based on human-computer interaction is mediated by computers or electronic devices. “Human-human” interaction, “human-machine” interaction, and “human-machine-human” interaction [10, 11] are all assisted by interaction.

Deep learning is a new research direction of machine learning. It learns tasks by simulating the neural network structure of the human brain [12]. Deep learning technology has gradually been applied to various fields and has been successfully applied to the related processing of pictures, audio, and video with the continuous advancement of science and technology [13]. The adoption of deep learning in the field of natural language processing started relatively late. Nevertheless, it solves many problems such as machine translation (Google translation), question-and-answer system (which can automatically generate corresponding feedback and answers based on user questions), and sentiment analysis (which can analyze the emotion of speech to generate relevant decisions). There are other functions of information extraction, tag sequence, and syntax analysis [14]. Machine learning techniques are currently widely used for related sentiment analysis of microblog, such as support vector machines, K-means clustering methods, and naive Bayes [15].

2.2 The adoption status of neural network in sentiment analysis

Research on application of deep learning in microblog sentiment analysis is focused on CNNs and recurrent neural networks only. The disadvantage of these networks is the processing sequence of short text data. Most sentiment analysis of Chinese microblogs can only divide microblogs into positive and negative categories [16, 17]. Topal and Kurt [18] found that the main motivation behind hidden communication was hiding the existence of communication between two users. In this work, the covert communication performance in a non-orthogonal multiple access non-orthogonal multiple access (NOMA) network with amplifying and forwarding relay was studied. Li et al. [19] revealed that high-resolution remote sensing images can provide rich feature information in the task of change detection. However, the interference of noise and complex background information can also bring some challenges to CD. In recent years, deep learning methods represented by CNN have harvested good results.

3. Materials and methods

3.1 Sentiment analysis of reposting microblog under LDA model

I. Establishment of microblog sentiment dictionary. The first step is text preprocessing. Institute of Computing Technology, Chinese Lexical Analysis System (ICTCLAS) word segmentation system [20] can support users to customize the dictionary. Basic emotional words and the words contained in the online emotional dictionary are added to the ICTCLAS word segmentation system. Then, microblog words and sentences are guided into the word segmentation system for data set processing. A stop word dictionary is established based on the characteristics of the microblog text, and the stop words in the word segmentation result are deleted. Second, the emotional benchmark words in Chinese vocabulary and microblog words are chosen. Half of the words are commendatory, and the other half are derogatory (praise: ha-ha, beautiful, nice, happy, perfect, praise, etc.; derogatory sense: garbage, tragedy, horror, regret, spit, etc.). Third, emotional candidate words are selected from microblog words. Template level words are adopted to select emotional candidate words for microblog (savory, powerful, abusive, beautiful, good future, like you, true love, very unfair, high, unaffordable, etc.). Fourth, the emotional trend of the emotional candidate words is judged, and the number of times the two words appear together are calculated to obtain the Point Mutual Information (PMI) [21] value, as shown in Eq. (1).

$\displaystyle\textit{PMI}({A,B})=\log_{2}\frac{P({A\&B})}{P(A)P(B)}$ (1)

In Eq. (1), $P(A)$ and $P(B)$ represent the probability of words $A$ and $B$ appearing separately, and $P(A{\&}B)$ represents the probability of $A$ and $B$ colleagues appearing. The PMI value represents the correlation between words.

The emotional tendency of words can be calculated from the PMI value, as shown in Eq. (2).

$\displaystyle C-\textit{PMI}(N)=\sum{{}_{K_{m}\in K}}\textit{PMI}({N,K_{m}})-% \sum{{}_{j_{m}\in K}}\textit{PMI}({N,j_{m}})$ (2)

In Eq. (2), $N$ represents microblog words, $K$ represents a vocabulary set containing positive emotional tendencies, and $J$ represents a vocabulary set containing negative emotional tendencies. If C-PMI $\geqslant$ 0, then microblog words have a positive emotional trend; but if C-PMI $<$ 0, they have a negative emotional trend.

III. Establishment of LDA model. LDA model [22] generates probabilistic processing discrete data (corpus). The LDA model is a topic model, using probability distribution of the topic to build a model for each document in the document set (Fig. 1).

Figure 1.

LDA model.

The box on the right side of the Fig. 1 represents the topic, and $E$ represents the number of topics. The outer box on the left represents documents, $H$ represents the number of documents. The inner dashed ellipse represents the selection of topics and words in the document, and $O$ represents the number of words in the document. LDA set $\alpha$ as the prior hyperparameter of the subject in the document set, and $X$ as the term prior hyperparameter in the subject set. Each document obeys the Dirichlet distribution with parameter as $\alpha$ [23], and each topic obey the Dirichlet distribution with parameter as $X$ . $\alpha$ represents the strength of the hidden topics of the document, and $X$ represents the probability distribution of all hidden topics. To specify a document set, LDA proposes the topic $V$ implicit in the document by the Eq. (3) of the posterior probability.

$\displaystyle P({V|R})=\frac{P({R,V})}{\sum{{}_{V}P({R,V})}}$ (3)

After all the microblog content of a certain user are trained, the blog content theme of the user is modelled. User characteristics, microblog characteristics, and content theme characteristics (topic vectors) are combined to form a microblog forwarding behavior characteristic model. Thus, a list of forwarding characteristics is formed, as shown in the Table 1.

I. Sentiment analysis of microblog forwarding.

The sentiment value of microblog $U$ is set as $G U$ , the sentiment value of the forwarding sentence $D$ is $G_{z}$ , and the calculation for sentiment analysis of microblog forwarding behavior is shown in Eq. (4).

$\displaystyle s=\sum\nolimits_{n=1}^{f,l}{G_{U}}\cdot G_{Z}\cdot V\cdot A\cdot% \textit{sim}({f,u})$ (4)

In Eq. (4), $A$ represents the relationship influence weight, and $\textit{sim}(f,u)$ represents the similarity of users’ perception of microblog.

Table 1

Analysis of the status quo of advanced research

Year	Literature title	Advantages	Disadvantages
2015	[10] Vision based hand gesture recognition for human computer interaction: a survey.	The existing literature related to gesture recognition systems for human-computer interaction is analyzed by classifying them under different key parameters.	The accuracy of human-computer interaction gesture recognition needs to be enhanced.
2016	[11] Semantics-based Intelligent Human-Computer Interaction.	A framework is proposed to automatically configure the behavior of the system based on user input and contextual information.	When the workload and complexity related to the interactive logic configuration increase, performance needs to be verified.
2016	[12] Deep Learning.	Summary of research results of deep learning technology in various fields is made in this work.	Deep learning can be further explored.
2016	[14] Image matting in the perception granular deep learning.	A generalized matting method is defined in the novel granular deep learning framework.	Effectiveness can also be enhanced in foreground and background separation tasks.
2018	[9] Technology: Artificial intelligence.	Incorporating artificial intelligence into the teaching and learning process can greatly improve the way students perceive knowledge.	Meaningful and differentiated course directions for AI design need to be studied.
2018	[13] Artificial Intelligence in the Rising Wave of Deep Learning: The Historical Path and Future Outlook [Perspectives].	It proposes and discusses how machine learning uses a set of algorithms to solve problems on its own.	Algorithm research is not deep enough.
2018	[16] Microblog sentiment analysis method based on a double attention model.	The dual attention model is used for microblog sentiment analysis, and the microblog sentiment symbol knowledge base is constructed based on the existing sentimental words, degree adverbs, negative words, microblog expressions and Internet common words, and other emotional semantic resources.	The accuracy of perception models can be improved.
2020	[17] A content search method for security topics in microblog based on deep reinforcement learning.	The microblog search for specific topics is modeled as a continuous state Markov decision process, and a security topic microblog search method based on deep reinforcement learning is proposed.	Search for complex topics needs to be studied.
2020	[18] Covert Communication in Cooperative NOMA Networks//2020 28th Signal Processing and Communications Applications Conference (SIU).	The performance of covert communication in non-orthogonal multiple access NOMA networks with amplifying and forwarding relay is studied.	Performance needs to be improved.
2021	[19] A Combined Loss-Based Multiscale Fully Convolutional Network for High-Resolution Remote Sensing Image Change Detection.	A multi-scale full convolutional neural network (MWNN) based approach is proposed, which uses multi-scale convolutional kernel to extract the detailed features of ground features.	The influence of noise and complex background information on image changes should be taken into account.

3.2 Deep learning sentiment classification model

3.2.1 LSTM sentiment classification model

LSTM [24] is an optimized network belonging to the recurrent neural network, which updates the current node information in the form of three gates, so as to effectively memorize historical information. It calculates the next node according to the values of the previous node and the current node. The LSTM sentiment classification model includes the following three parts.

Input layer. The Word2vec tool is used to train Chinese microblog texts with well-divided words, and the word vectors containing contextual semantic information become the input vocabulary of the LSTM model.

Hidden layer. This layer performs feature learning. A vector representation of the word is obtained after the first training. Sentences in microblog are regarded as a collection of word vectors. The overall vector of the sentence is obtained as the input part of the LSTM model after the word vectors are concatenated. The LSTM model can fully and effectively use the context information of the text for feature learning. Under the action of the three gates, the state value of the hidden layer is changed according to the historical information of the input sequence, to obtain the feature vector.

Output layer. As the last layer of the model, it uses the activation function to calculate the classification probability. Sigmoid function [25] is used, after the hidden layer of the model completes feature learning, it is connected to the Sigmoid layer, and the classification probability is calculated through the Sigmoid function. From the calculation result, the sentiment classification result of the microblog is obtained, and the accuracy of the sentiment classification of the microblog is output. The model calculation process is shown in Fig. 2.

Figure 2.

LSTM sentiment classification model calculation process.

3.2.2 CNN sentiment classification model

Initially, the core application of the CNN sentiment analysis model [26] is image processing. It is now possible to apply natural language processing to the principles of CNN image processing. The first two data of the array in the expression of a sentence indicate the length of the sentence and the dimension of the word vector in turn. The proposed CNN sentiment classification model mainly consists of four parts.

The input layer uses word2vec model training to obtain word vectors representing words of text and uses the text as the input part of the CNN model. The convolution layer belongs to the feature learning layer. Several convolution kernels of different sizes are used to learn sentence features, and several feature expressions are obtained. The maximum pooling in the pooling layer compresses the features learned by the convolutional layer to obtain the core features. The fully connected layer connects all the features obtained by the collection layer, and then inputs the obtained feature vector into the classifier.

4. Results

4.1 Experimental data sets

The data of this research is taken from Tencent Weibo, and Application Programming Interface (API) crawler method [27] is adopted to obtain the microblog background data set. 41,235 microblogs from June 1–28, 2021 are selected as training data samples, and 4,126 microblogs from June 29–31 as test data samples. From April to June, the number of microblogs was 329,876, an epic number. To obtain accurate sentiment classification results, artificial sentiment labeling measures are adopted for the content of 4,100 reposted comments in the test data. After text preprocessing, 3,125 pieces of data are obtained. Of them, 1,003 items belong to positive emotional tendencies, and 406 items belong to negative emotional tendencies. Others are 1,716 items (positive emotions include happiness, encouragement, like, and worship, while negative emotions include hatred, anger, hatred, and sadness). Positive emotion prediction is defined as $M N$ , and negative emotion prediction is defined as $F N$ . The negative prediction is defined as the positive category as $F P$ . Then, the calculations of the evaluation indexes used for sentiment analysis, precision (P’), the recall rate (R), and the F1 value, are shown in Eq. (4.1).

$\displaystyle P^{\prime}=\frac{MN}{MN+FN}$ $\displaystyle R=\frac{MN}{MN+FP}$ (5) $\displaystyle F1=\frac{2*P^{\prime}*R}{R+P^{\prime}}$

4.2 Experimental environment configuration of deep learning mode

The text classification experiment of the deep learning model is completed by the deep learning framework Keras [28]. The backend of Keras is Tensorflow or Theano. The deep learning lib library is a deep learning library that contains many currently very commonly used learning models, such as CNN and RNN. Keras is an advanced neural network that will include TensorFlow. Keras supports CNN and RNN (or a combination of the two), completes quickly and concisely, and can quickly transform ideas into experimental results. Moreover, it supports switching between CPU and GPU at any time. The design principles of Keras are as follows.

1.
Keras provides a concise API, and the code implementation process is relatively simple, which reduces the workload and improves work efficiency.
2.
The network layer, activation function, loss function, and optimizer in the neural network are all encapsulated in different areas. High encapsulation allows users to arbitrarily combine modules according to actual needs to build their own models.
3.
Keras imitates the current module, and the new function written can be extended and innovated in the module.
4.
The models in Keras are all written in Python code [29], which is very scalable and convenient.

The hardware and software configuration of the experimental environment is shown in Table 3.

Table 2
Forwarding feature list

User characteristics Number of fans; degree of attention; number of microblogs; whether or not there is certification;

level; number of microblogs transferred

Microblog features Are there any pictures or videos; number of reposts; number of comments; number of likes

Table 3
The hardware and software configuration of the experiment

Lab environment Environment configuration

Operating system Windows 10

CPU Inter(R) Core (TM)i5-9400 CPU @2.90 GHz

GPU NVIDIA GeForce GTX I660 Ti

PYTHON Python 3.7.6

PYTORCH-GPU 1.2.0

4.3 Data preprocessing

Lab environment	Environment configuration
Operating system	Windows 10
CPU	Inter(R) Core (TM)i5-9400 CPU @2.90 GHz
GPU	NVIDIA GeForce GTX I660 Ti
PYTHON	Python 3.7.6
PYTORCH-GPU	1.2.0

The content of the original data set obtained by the API crawler is very messy, which also contains useless symbols or characters. It is necessary to format the obtained data before data analysis. The main content of data preprocessing is data cleaning and Chinese word segmentation.

The first step is the data cleaning. Redundant and messy data are reduced, and the emotional analysis of the data that is meaningless in the later stage is also reduced, so that the efficiency and accuracy of the experiment is improved. The main methods of data cleaning are as follows.

1.
Delete useless symbols in the data. Some bloggers will add symbols to words because of personal habits when they post on microblog, and these symbols are not common symbols, which have no practical significance for the later sentiment analysis, thus are removed.
2.
Delete the URL in the data. Some microblogs have a URL attached to the body content, which can be clicked on to where this microblog comes from. These URLs have no meaning for analyzing the sentimental issues of reposting microblog, so the URLs should be deleted.

Second, word segmentation measures are not needed, for Chinese word segmentation and English words are separated by spaces. However, the structure and grammar of Chinese are not very complicated. The same sentence may have multiple meanings and different combinations of words, which will increase the difficulty of Chinese word segmentation. The accuracy of word segmentation determines the accuracy of sentiment analysis results.

Currently, the most popular word segmentation tools are BosonNLP, language cloud, and Natural Language Processing & Information Retrieval (NLPIR) [30, 31]. The wording of microblog data is very serious, the types of words used are very diverse and casual, and the emerging vocabulary on the Internet is often used. Due to the characteristics of microblog data, BosonNLP [32] is very helpful to improve the accuracy of Chinese microblog word segmentation according to the starting point of new word recognition and the perspective of combining different meanings. BosonNLP is adopted to process and analyze Chinese word segmentation in the experimental data of microblog.
4.4 Microblog text experimental classification results

The sentence text will be supplemented with massive meaningless data if the number of qualifiers in the microblog text is too large, which increases the amount of calculation and affects the model’s extraction of semantic features. If the number of qualifiers is too small, the rich semantic information in the text will be wasted. Therefore, the number of qualifiers needs to be determined through experiments. Before the experiment, the frequency distribution of the microblog text length of the data set is drawn as shown in Fig. 4.

Figure 3.

CNN sentiment classification process.

Figure 4.

Microblog text length distribution.

In Fig. 4, the text length is mostly concentrated below 80 words, so it is relatively appropriate to select about 80 words as the number of experimental words. In this experiment, the length of the text is 10 words as the interval, a total of 5 words from 60 words to 100 words are tested, and their accuracy and F1 value on the test set are compared.

It is impossible to put the entire training set data into the model for training at one time because of the very large sample size of the data set for deep neural network training. The training set is usually divided into multiple mini-batches of the same number in practical applications. The model receives a batch of data for training each time, compares the output results obtained with the actual results of the batch, calculates the loss function, and updates the model parameters through backpropagation.

In general, the batch size should be kept at a reasonable level. The direction is random if the batch is small, though the model parameters are updated frequently, and it may be difficult to converge. It may be difficult to achieve due to the performance of the computer if the batch is large, though the direction of model training can be guaranteed to be stable. In this experiment, a total of 6 batch sizes of 16, 32, 64, 128, 256, and all data sets are used for comparison experiments. The training time of the model and the evaluation indicators on the test set are recorded. The results are shown in Table 4.

Table 4

Comparison results of batch size experiments

Batch size	Time spent (s)	Accuracy (%)	F1 (%)
16	740	91.23	91.45
32	470	91.34	91.35
64	311	91.76	91.56
128	228	92.08	91.88
256	197	91.66	91.57

As the sample size of a single batch increases, the time to run ten Epochs gets shorter and shorter. However, the GPU memory can’t be trained because it is full, if the batch size is the entire data set. From the evaluation index, the F1 value and accuracy do not tend to be ideal with the increase of the training batch size. The classification performance of the model is the best when the batch size is 128.

4.5 Comparison of feature representations of linear discriminant analysis models

With and without the introduction of emotional features in LDA, the number of topics in the LDA model is set to 30, 50, 80, 110, 140, 170, and 200, and the sentiment value under different number of topics is analyzed. The accuracy, recall, and F1 values are calculated in Fig. 5a and b.

When the LDA model does not introduce emotional features, if the number of topics is 120, F1 $=$ 82.9%, which is the maximum value. When the LDA model introduces emotional features, if the number of topics is 120, F1 $=$ 84.9%, which is the maximum value. The proposed LDA model is universal and can accurately analyze the sentiment of users reposting microblog.

4.6 Sentiment analysis results of user reposting microblog

I. Comparison of results of sentiment binary classification between CNN model and LSTM model.

Many parameters need to be adjusted in order not to affect the experimental results if CNN and LSTM models are employed to classify microblog sentiment. It is assumed that the maximum input length of a microblog sentence (Maxlen) is 130, the amount of data for each parameter update is (Batch _ size): 34, the word vector dimension (Embedding_Vecor_Length) is: 110, the number of data iterations (N_Epoch) during training is 15, and the ratio of the training set to the test set (Test_Size) is 0.3. The comparison results of the emotional binary classification results of the CNN model and the LSTM model are shown in Table 2.

Goodness of fit refers to the degree of fit of the regression line to the observed value. The statistic that measures the goodness of fit is the coefficient of determination (also known as the coefficient of determination) $R^{2}$ . The maximum value of $R^{2}$ is 1. The closer the value of $R^{2}$ is to 1, the better the fit of the regression line to the observed value. Conversely, the smaller the value of $R^{2}$ , the worse the fit of the regression line to the observed value.

Table 5
Comparison of the results of sentiment binary classification between CNN model and LSTM model

Model	Forecast accuracy	Actual output accuracy	$R^{2}$
CNN	0.786	0.793	0.89
LSTM	0.839	0.849	0.91

Table 6

Chinese corresponding to English abbreviations

No.	English abbreviations	English full names
1	LDA	Latent Dirichlet Allocation
2	CNN	Convolutional Neural Network
3	LSTM	Long Short-Term Memory Network
4	API	Application Programming Interface
5	PMI	Point Mutual Information
6	RNN	Recurrent Neural Network
7	CPU	Central Processing Unit
8	GPU	Graphics Processing Unit
9	URL	Uniform Resource Locator
10	NLPIR	Natural Language Processing & Information Retrieval

Figure 5.

LDA model feature representation. a: no emotional features are introduced; b: emotional features are introduced.

Table 7

Equation symbols

No.	Symbols	Meanings
1	P (A)	Probability of word A
2	P (B)	Probability of word B
3	P (A&B)	Probability of A and B appearing at the same time
4	N	Microblog words
5	K	Vocabulary set containing positive affective tendencies
6	J	Vocabulary set containing negative affective tendencies
7	X	Priori parameters of terms in the subject set
8	$\alpha$	Strength of document hiding topics
9	E	Number of topics
10	V	Implied topics in the document
11	a	Relationship influence weight
12	sim (f, u)	The similarity of users’ perception of microblog
13	MN	Positive emotion prediction
14	FN	Negative emotion prediction

Table 8

Pseudo code of microblog sentiment analysis algorithm

1.	Start
2.	Input: N, K, J, P (A), P (B)
3.	Parameters: $n=$ microblog words, $k=$ vocabulary set containing positive emotions,
4.	$j=$ vocabulary set containing negative emotions
5.	P (A) $=$ probability of occurrence of word A, P (B) $=$ probability of occurrence of word B.
6.	The microblog text is preprocessed by using (ICTCLAS) word segmentation system.
7.	If the word has been deactivated, add it to the deactivation dictionary
8.	The emotional reference words in Chinese words and microblog words are divided into commendatory words and
	derogatory words
9.	Select emotional candidate words from microblog words
10.	Judge the emotional trend of emotional words and calculate the probability that two words a and B appear together:
	$\textit{PMI}({A,B})=\log_{2}\frac{P({A\&B})}{P(A)P(B)}$
11.	Calculate the tendency of words: $C-\textit{PMI}(N)=\sum{{}_{K_{m}\in K}}\textit{PMI}({N,K_{m}})-\sum{{}_{j_{m}% \in K}}\textit{PMI}({N,j_{m}})$
12.	If C-PMI $\geqslant$ 0, microblog words have a positive emotional trend
13.	If C-PMI $<$ 0, microblog words have a positive emotional trend
14.	Output: Emotional tendency of forwarding microblog
15.	End

Figure 6.

ROC curve of CNN and LSTM sentiment analysis model.

From Table 5, the actual classification accuracy of the CNN model is 0.793, and that of the LSTM model is 0.849. Both have high accuracy in microblog sentiment classification, and the LSTM sentiment classification model has a fitting degree of 0.91, which is close to 1, indicating that the classification is accurate and LSTM is suitable for users to forward microblog sentiment classification.

Figure 7.

Sentiment analysis of microblog forwarding under different feature combinations. a: positive trend; b: negative trend.

Figure 8.

Performance comparison of different sentiment analysis methods.

Figure 9.

Flow chart of the overall research framework.

Area under curve (AUC) is defined as the area under the receiver operating characteristic (ROC) curve and the coordinate axis. Obviously, the value of this area will not be greater than 1. Since the ROC curve is generally above the line $y=x$ , the value of AUC ranges between 0.5 and 1. The closer the AUC is to 1.0, the higher the authenticity of the detection method. When it is equal to 0.5, the authenticity is the lowest and it has no application value. The criteria for judging the pros and cons of a classifier (predictive model) from AUC are as follows.

AUC $=$ 1, which means the perfect classifier.

AUC $=$ [0.85, 0.95], which means the effect is good.

AUC $=$ [0.7, 0.85], which means the effect is general.

AUC $=$ [0.5, 0.7], which means that the effect is low, but it is already very good for predicting stocks.

AUC $=$ 0.5, which means that it likes random guessing (for example: losing a copper plate), and the model has no predictive value.

AUC $<$ 0.5, which means that it’s worse than random guessing, but as long as it always goes against prediction, it’s better than random guessing.

Table 9

Comments and responses

No.	Comments	Sales note	Reply	Edit note
1	Reviewer #2: Please see the attached file. This manuscript studied the users’ sentiment from their reposting Weibo data by using the latent dirichlet allocation model in the framework of deep learning. This work is an interesting study. However, some other problems in the manuscript are still concerned in the following: 1. In the experiments, could the authors make a comparison with more state-of-the-art methods to validate the effectiviness? 2. The contribution should be clarified. 3. The organization of this manuscript should be added to the end of the introduction. 4. The format of mathematical equations is not satisfactory. 5. I suggest the authors to show a flow chart of this work to improve the readability for the future readers. 6. More related works on convolutional neural network should be included in the introduction, such as “DOI: 10.1109/SIU49456.2020.9302448”, “DOI: 10.1109/LGRS.2021.3098774” …7. Please check Ref. [11]. It seems that the journal is wrong.		Reviewer #2: Please see the attached file. This manuscript studied the users’ sentiment from their reposting Weibo data by using the latent dirichlet allocation model in the framework of deep learning. This work is an interesting study. However, some other problems in the manuscript are still concerned in the following: Thank you for your affirmation. I will revise the questions carefully. 1. In the experiments, could the authors make a comparison with more state-of-the-art methods to validate the effectiviness? Thank you for your valuable comments. This research method has been added in the results, compared with advanced research in other related fields, and the comparison curve is given. See section 4.4. 2. The contribution should be clarified. Thank you for your valuable comments. The abstract and introduction of the article have been revised to highlight the specific contributions of the article. 3. The organization of this manuscript should be added to the end of the introduction. Thank you for your valuable comments. The overall structure of the article has been added to the end of the introduction. 4. The format of mathematical equations is not satisfactory. Thank you for your suggestion. The format of the equation has been modified. 5. I suggest the authors to show a flow chart of this work to improve the readability for the future readers. Thank you for your comments. The overall flow chart of this study has been added at the end of the article. See Fig. 9. 6. More related works on convolutional neural network should be included in the introduction, such as “DOI: 10.1109/SIU49456.2020.9302448”, “DOI: 10.1109/LGRS.2021.3098774” …Thank you for your comments. The references and content introduction of convolutional neural network have been added in the introduction, and the cited references have been added in the list of references. See references [18, 19]. 7. Please check Ref. [11]. It seems that the journal is wrong Thank you for your comments. Reference [11] has been modified. Now it is literature [8].
2	Reviewer #3: Journal: Soft Computing Manuscript Title: Sentiment Analysis of Users Reposting Weibo under Latent Dirichlet Allocation Deep Learning Model Manuscript ID: SOCO-D-21-01696 Reviewer Number: 3 Submission Date: Monday, August 16, 2021		Reviewer #3: Journal: Soft Computing Manuscript Title: Sentiment Analysis of Users Reposting Weibo under Latent Dirichlet Allocation Deep Learning Model Manuscript ID: SOCO-D-21-01696 Reviewer Number: 3 Submission Date: Monday, August 16, 2021 The manuscript presented “a sentiment analysis approach using deep learning”.

Table 9, continued
No.	Comments	Sales note	Reply	Edit note
	The manuscript presented “a sentiment analysis approach using deep learning”. However, the major and critical weak points are that: (1) Their proposed work discussion is weak distributed to be described or analyzed. (2) The novelty is not guaranteed. (3) Their work is not compared with state-of-the-art approaches nor related studies. (4) Their experiments leak from the descriptive and statistical analysis.		However, the major and critical weak points are that: (1) Their proposed work discussion is weak distributed to be described or analyzed. Thank you for your advice. The discussion has been revised and added in the results section. See 4. Results. (2) The novelty is not guaranteed. Thank you for your comments. As the most cutting-edge computer algorithm, deep learning has unique advantages. In this paper, convolutional neural network and long-term and short-term memory network algorithm are applied to emotion analysis, which itself is a novel idea. (3) Their work is not compared with state-of-the-art approaches nor related studies. Thank you for your valuable comments. This research method has been added in the results, compared with advanced research in other related fields, and the comparison curve is given. See section 4.4. (4) Their experiments leak from the descriptive and statistical analysis. Thank you for your valuable comments. The goodness of fit analysis of predicted results and actual output results has been added in the results section. See 4. Section 3.
3	The rest of my review presents other weak points, comments, and opinions in detail. Overall Comments: (1) The keyword “deep learning” in the title is common. Please, be consice. The authors is recommended to rename it. (2) [KEYWORDS] The keywords should be sorted in alphabetical order. (3) [ABSTRACT] The abstract should contain the best-achieved results from the performed experiments. (4) [ABSTRACT] The abstract should reflect the contributions of the manuscript. I suggest rewriting it. (5) [INTRODUCTION] The authors should provide a clear problem definition and contributions in the introduction section. (6) [RELATED WORK] Where are the related studies? They should be declared in a separate section. (7) [RELATED WORK] A table of comparisons should be added at the end of the related studies section to praise the pros. and cons. of them. The year column should be added and they should be ordered by it. (8) [EQUATIONS] The authors should follow the journal authors’ guidance in writing the equations, symbols, and variables.		The rest of my review presents other weak points, comments, and opinions in detail. Overall Comments: (1) The keyword “deep learning” in the title is common. Please, be consice. The authors are recommended to rename it. Thank you for your comments. The “deep learning” in the title has been modified to the refined technology “neural network”. (2) [KEYWORDS] The keywords should be sorted in alphabetical order. Thank you for your comments. The keywords have been rearranged alphabetically. (3) [ABSTRACT] The abstract should contain the best-achieved results from the performed experiments. Thank you for your comments. The best results of the experiment have been added to the summary. (4) [ABSTRACT] The abstract should reflect the contributions of the manuscript. I suggest rewriting it. Thank you for your valuable comments. The abstract has been rewritten in the format of “research purpose – research methods – research results – research conclusions” to highlight the contribution of this study. See summary. (5) [INTRODUCTION] The authors should provide a clear problem definition and contributions in the introduction section. Thank you for your comments. The overall structure of the introduction has been adjusted to highlight the problems existing in the current research and the outstanding contributions of this research. See 1. Introduction. (6) [RELATED WORK] Where are the related studies? They should be declared in a separate section. Thank you for your comments. For the research on advanced technologies in related

Table 9, continued
No.	Comments	Sales note	Reply	Edit note
	(9) [METHODOLOGY] The suggested approach is not clearly discussed. More details should be added. (10) [METHODOLOGY] Where are the used equations of the used metrics? (11) [METHODOLOGY] What are the used equations in the suggested approach? (12) [METHODOLOGY] Where is the overall pseudocode? Flowchart? of the suggested approach? (13) [DATASETS] Samples from the used dataset should be added and annotated. (14) [EXPERIMENTS] The experimental configurations should be added to a table. (15) [EXPERIMENTS] What are the criteria for selecting the experimental configurations? (16) [EXPERIMENTS] More experiments should be conducted using different configurations. (17) [EXPERIMENTS] Where is the tabular representation of the reported results? (18) [EXPERIMENTS] The figures in the experiments section should be gridded. (19) [EXPERIMENTS] Why did not the authors compare their approach with others in a table? (20) [EXPERIMENTS] Why did not the authors compare their approach with another approach to compare the suggested approach efficiency and applicability? (21) [EXPERIMENTS] Can the authors draw the area under the curve (AUC)? (22) [EXPERIMENTS] Why did not the authors calculate other performance metrics such as specificity and f1-score? (23) [EXPERIMENTS] Where is the detailed and statistical discussion of the reported results? (24) [EXPERIMENTS] More experiments should be conducted using a different dataset to prove the generalization. (25) [ABBREVIATIONS] The authors should add a table of abbreviations in the revised manuscript. (26) [SYMBOLS] The authors should add a table of symbols in the revised manuscript. (27) [CONCLUSIONS] The conclusions in this		fields, the introduction is modified, the relevant research is listed as a literature review, and the contents of large paragraphs are used to summarize and describe the previous research results. See 2. Literature review. (7) [RELATED WORK] A table of comparisons should be added at the end of the related studies section to praise the pros. and cons. of them. The year column should be added and they should be ordered by it. Thank you for your comments. A literature review has been added in the second part, and a list of current advanced research has been added at the end of the paragraph, including the advantages and disadvantages of the research. See Table 1. (8) [EQUATIONS] The authors should follow the journal authors’ guidance in writing the equations, symbols, and variables. Thank you for your comments. The equations in the text have been modified and checked, and the symbols and variables have been explained accordingly. (9) [METHODOLOGY] The suggested approach is not clearly discussed. More details should be added. Thank you for your comments. The method part has not been pointed out in detail, and the frame diagram has been supplemented. Model Figs 2 and 3 are supplemented. (10) [METHODOLOGY] Where are the used equations of the used metrics? Thank you for your comments. In Section 3.3, the expression of the evaluation index equation for the accuracy, recall (R) and F1 value used has been given. (11) [METHODOLOGY] What are the used equations in the suggested approach? Thank you for your comments. The corresponding specific equations have been given in the method. (12) [METHODOLOGY] Where is the overall pseudocode? Flowchart? of the suggested approach? Thank you for your question. A screenshot of the method pseudo code has been attached at the end of the article. The flow chart of the method is shown in Figs 1–3, which have been given in the corresponding method part. (13) [DATASETS] Samples from the used dataset should be added and annotated. Thank you for your comments. The data sets and samples used in this research experiment have been indicated in the text. See Section 3.3 for details. (14) [EXPERIMENTS] The experimental configurations should be added to a table.Thank you for your valuable comments. The experimental hardware and software configuration table has been added in the experimental environment configuration section of Section 3.4. (15) [EXPERIMENTS] What are the criteria for selecting the experimental configurations? Thank you for your question. The experimental configuration is configured according to the requirements of the software and hardware environment required for the operation of deep learning to ensure the normal operation of the model data.

Table 9, continued
No.	Comments	Sales note	Reply	Edit note
	manuscript are primitive. Please, write your conclusions. (28) [REFERENCES] There are no citations for many sentences in the manuscript. Why? Please check. (29) [REFERENCES] The references should be written in the same style following the journal authors’ guidance. (30) [REFERENCES] Recent citations from 2021 should be added to the manuscript. (31) [PROOFING] The authors should get editing help from someone with full professional proficiency in English. (32) [CONSISTENCY] The manuscript structure is too short. It must be elaborated in their applied technology as should support more rigorous technical aspects. (33) [CONSISTENCY] Some paragraphs are wrapped in more than 10 lines. They should be split concisely. (34) [NOVELTY] What is the novelty of the suggested approach? (35) [FIGURES] The authors should provide high-resolution figures in the manuscript. For example, Fig. 2. (36) [LIMITATIONS] What are the limitations of the current study? It should be added in a separate section. (37) Please, add a page number at the bottom of the page. (38) Figure 2 is not clear. (39) Where is the deep learning model in the study? Section 2.2 does not illustrate it. For the authors in case of the authors got a chance to review the manuscript and submit the revised one after the editor’s decision, please, provide a table in the revised manuscript mentioning (1) the comment, (2) the authors’ response, and (3) the authors’ change (if applicable). Please, consider all of the comments and don’t ignore any of them. Please, refer to the attached file “SOCO-D-21-01696 Reviewer.pdf” for the same comments in an organized format.		(16) [EXPERIMENTS] More experiments should be conducted using different configurations. Thank you for your proposal. The experimental environment configuration is the optimal configuration selected according to the requirements of the software and hardware environment required in this study. If the experiment is replaced with other configurations, the classification accuracy of the results will be affected. Therefore, no more configurations are required. (17) [EXPERIMENTS] Where is the tabular representation of the reported results? Thank you for your comments. The data chart of experimental results has been added in the results Section 4.1. (18) [EXPERIMENTS] The figures in the experiments section should be gridded. Thank you for your comments. The pictures in the article have been modified and the grid has been added. See all curves in the text. (19) [EXPERIMENTS] Why did not the authors compare their approach with others in a table? Thank you for your valuable comments. This research method has been added in the results, compared with advanced research in other related fields, and the comparison curve is given. See section 4.4. (20) [EXPERIMENTS] Why did not the authors compare their approach with another approach to compare the suggested approach efficiency and applicability? Thank you for your valuable comments. This research method has been added in the results, compared with advanced research in other related fields, and the comparison curve is given. See section 4.4. (21) [EXPERIMENTS] Can the authors draw the area under the curve (AUC)? Thank you for your comments. The ROC curves of two emotion analysis models have been given in Section 4.3 of the results, and the area under the curve (AUC) has been calculated. (22) [EXPERIMENTS] Why did not the authors calculate other performance metrics such as specificity and f1-score? Thank you for your comments. The performance index of the experimental results has been modified and the specificity and F1 value calculation have been added. See the results section. (23) [EXPERIMENTS] Where is the detailed and statistical discussion of the reported results? Thank you for your comments. The experimental results of each part have been discussed in detail in the results section. See the results section. (24) [EXPERIMENTS] More experiments should be conducted using a different dataset to prove the generalization. Thank you for your proposal. This paper uses the data set collected by Tencent microblog. Tencent microblog itself is universal, and users account for more than half of all microblog users. Therefore, the collected data set is enough for experimental

Table 9, continued
No.	Comments	Sales note	Reply	Edit note
			analysis. There is no need to add data sets from other platforms. (25) [ABBREVIATIONS] The authors should add a table of abbreviations in the revised manuscript. Thank you for your comments. The abbreviation correspondence table has been added at the end of the text. See Table 6. (26) [SYMBOLS] The authors should add a table of symbols in the revised manuscript. Thank you for your comments. A symbol comparison table has been added at the end of the manuscript. See Table 7. (27) [CONCLUSIONS] The conclusions in this manuscript are primitive. Please, write your conclusions. Thank you for your comments. The conclusion has been modified. See 5. Conclusion. (28) [REFERENCES] There are no citations for many sentences in the manuscript. Why? Please check. Thank you for your comments. The part without quotation in the manuscript has been modified. See 1. Introduction and 2. Overview. (29) [REFERENCES] The references should be written in the same style following the journal authors’ guidance. Thank you for your comments. The format of references has been unified according to the requirements of journals. (30) [REFERENCES] Recent citations from 2021 should be added to the manuscript. Thank you for your comments. The latest citation of 2021 has been added at the end of the references. They are references [19, 33, 34, 35, 36]. (31) [PROOFING] The authors should get editing help from someone with full professional proficiency in English. Thank you for your proposal. The manuscript has been proofread and checked by professional translators whose native language is English. (32) [CONSISTENCY] The manuscript structure is too short. It must be elaborated in their applied technology as should support more rigorous technical aspects.Thank you for your suggestions. The structure of the manuscript has been adjusted, the length has been increased, and the technologies and methods used in the research have been improved. (33) [CONSISTENCY] Some paragraphs are wrapped in more than 10 lines. They should be split concisely. Thank you for your valuable comments. The long paragraphs in the article have been separated. (34) [NOVELTY] What is the novelty of the suggested approach? Thank you for your comments. The innovation of this study has been described in the last part of the introduction. (35) [FIGURES] The authors should provide high-resolution figures in the manuscript. For example, Fig. 2.

Table 9, continued
No.	Comments	Sales note	Reply	Edit note
			Thank you for your comments. Figures 2 and 3 have been readjusted to improve the picture resolution. See Figs 5 and 7. (36) [LIMITATIONS] What are the limitations of the current study? It should be added in a separate section. Thank you for your comments. A separate description of the limitations of this study has been added in the conclusion. See Section 5.2. (37) Please, add a page number at the bottom of the page. Thank you for your comments. The page number of the article has been added at the bottom of the page. (38) Figure 2 is not clear. Thank you for your comments. Figure 2 has been readjusted to improve the resolution. See Fig. 5. (39) Where is the deep learning model in the study? Section 2.2 does not illustrate it. Thank you for your valuable comments. The structure diagrams of LSTM emotion analysis model and CNN emotion analysis model have been added in Section 2.2. Now is the revised section 3.2. For the authors in case of the authors got a chance to review the manuscript and submit the revised one after the editor’s decision, please, provide a table in the revised manuscript mentioning (1) the comment, (2) the authors’ response, and (3) the authors’ change (if applicable). Please, consider all of the comments and don’t ignore any of them. Thank you for your comments. A comparison table of comments and replies has been added at the end of the article. See Table 8. Please, refer to the attached file “SOCO-D-21-01696 Reviewer.pdf” for the same comments in an organized format. Thank you. I’ll refer to it.

Then, the ROC curve of CNN and LSTM models to classify microblog sentiment is shown in Fig. 6.

From Fig. 6, the AUCs of the CNN and LSTM sentiment analysis models are 0.85 and 0.91, respectively. The performance of the LSTM sentiment analysis model is better.

II. Sentiment analysis of microblog forwarding under different feature combinations.

The sentiment analysis of users forwarding microblog is carried out under three characteristics of user characteristics, user characteristics $+$ microblog characteristics, usage characteristics $+$ microblog characteristics $+$ relationship characteristics. The results are shown in Fig. 7a and b.

Adding microblog features can improve the experimental results, indicating that analyzing microblog users can increase topic interest. Since microblog users repost microblog from their own hobbies, microblog of similar nature also has similar emotional tendencies. Therefore, the polarity of reposted comment content has not changed much. The positive F1 value is 69.98%, the negative F1 value is 66.01% after the sentiment analysis of users and microblog features, and the difference is relatively small.

4.7 Comparison of sentiment analysis methods in this work with research in related fields

To prove the superiority of the sentiment analysis method proposed in this study, the proposed LDA sentiment analysis method and the sentiment analysis method based on CNN and LSTM are compared with the advanced ones proposed in references [33, 34, 35]. Their performances are compared regarding the Accuracy, Precision, Recall, and F1 values. The results are shown in Fig. 8.

From Fig. 8, the sentiment analysis method of this study and other current advanced research methods in the literature are compared and analyzed in Accuracy, Precision, Recall, and F1 values. The sentiment analysis model based on the LSTM network has a good classification accuracy, reaching 83.4%, and the values in the three directions of Precision, Recall, and F1 are also the highest. The LSTM sentiment analysis model has superior performance in the sentiment analysis process of users forwarding microblog.

5. Conclusion

5.1 Research summary

To analyze the sentiment attributes of users who forwarded microblog, a microblog sentiment dictionary is firstly established, then an LDA model is put forward. A microblog sentiment analysis model under deep learning, including CNN and LSTM models, is proposed. The LDA model, the CNN model, and the LSTM model are used to conduct sentiment analysis on users’ reposted microblogs, and the sentiment analysis results of the CNN model and the LSTM model are compared and analyzed. In the LDA model, sentiment analysis is carried out on microblog forwarding under different feature combinations. The results show that the LSTM model under deep learning is more suitable for sentiment analysis of microblog users. The LDA model can perform sentiment analysis on users’ reposting microblog based on the different characteristics of users and microblog and the relationship between the two. The accuracy of the two models is very high, which can promote the practical application of microblog in human-computer interaction education.

5.2 Limitations of this research

Several deep neural network sentiment analysis models are fabricated in this study. Some achievements are achieved in this study, which proves that the proposed CNN and LSTM sentiment analysis models can achieve good classification results. Limited by experimental conditions and technical level, however, there are still some deficiencies, which should be further improved.

The emotion dictionary established in this work is not comprehensive enough. It only consists of negative and positive emotions. In fact, human emotions are complex and changeable Negative emotions are also classified into sadness, regret, regret, and other different degrees of emotional expression. Therefore, more delicate emotion should be added to this text, which will be more conducive to the realization of the control of the emotional tendency of the text.

There is room for further mining of microblog comment information. This work only carries on the sentiment analysis to the microblog text. In fact, in addition to the Chinese text, pictures and videos may also reflect the reviewer’s emotional attitude, Therefore, multi-modal sentiment analysis of information content such as text and video in microblog may be better than single text sentiment analysis to grasp the emotional tendency of commentators.

Footnotes

Funding

This work was supported by following fundings:

Key Research Project of Guangdong Baiyun College, No. 2022BYKYZ02.

Key Research Platform of Guangdong Province, No. 2022GCZX009.

Special project in key fields of colleges and universities in Guangdong province, No. 2020ZDZX3009.

References

Cao

Lin

. A cross-media public sentiment analysis system for microblog. Multimedia Systems. 2016; 22(4): 479-486.

Svalova

Malysheva

Kozitsina

. Structure of the receptor layer in electrochemical immunosensors. Modern trends and prospects of development. Russian Chemical Bulletin. 2017; 66(10): 1797-1811.

Stark

Avitsur

Hunzeker

Padgett

Sheridan

. Interleukin-6 and the development of social disruption-induced glucocorticoid resistance. Journal of Neuroimmunology. 2016; 124(1-2): 9-15.

Pérez-Tellez

Cardiff

Rosso

Pinto

. Disambiguating company names in microblog text using clustering for online reputation management. Revista Signos. 2015; 48(87): 54-77.

Den Harder

Frijlingh

Ravesloot,

Oosterbaan

Gijp

AVD

. The importance of human-computer interaction in radiology e-learning. Journal of Digital Imaging. 2016; 29(2): 195-205.

Van Dalen

Van Rees

. More than just noise? Examining the information content of stock microblogs on financial markets. Journal of Information Technology. 2018; 33(1): 50-69.

Feng

Yingxue

. Research on China’s city network based on users’ friend relationships in online social networks: A case study of Sina Weibo. Geojournal. 2016; 81(6): 1-10.

Desire

Turner

D’Eath

Doeschl-Wilson

Lewis

CRG

Roehe

. Analysis of the phenotypic link between behavioural traits at mixing and increased long-term social stability in group-housed pigs. Applied Animal Behaviour Science. 2015; 166: 52-62.

Majumdar

Sarode

Patil

. Technology: Artificial intelligence. British Dental Journal. 2018; 224(12): 916.

10.

Rautaray

Agrawal

. Vision based hand gesture recognition for human computer interaction: A survey. Artificial Intelligence Review. 2015; 43(1): 1-54.

11.

Gatteschi

Lamberti

Montuschi

Sanna

. Semantics-based intelligent human-computer interaction. IEEE Intelligent Systems. 2016; 31(4): 11-21.

12.

Hao

Zhang

. Deep learning. International Journal of Semantic Computing. 2016; 10(03): 417-439.

13.

Deng

. Artificial intelligence in the rising wave of deep learning: The historical path and future outlook [perspectives]. IEEE Signal Processing Magazine. 2018; 35(1): 180-177.

14.

Pang

Shi

. Image matting in the perception granular deep learning. Knowledge-Based Systems. 2016; 102: 51-63.

15.

Miller

Mitchell

Barker

Giebel

Codling

Amory

Michie

Davison

Tachtatzis

Andonovic

Duthie

. Using animal-mounted sensor technology and machine learning to predict time-to-calving in beef and dairy cows. Animal. 2020; 14(6): 1-9.

16.

Zhang

Zheng

Huang

Jiang

. Microblog sentiment analysis method based on a double attention model. Journal of Tsinghua University. 2018; 58(2): 122-130.

17.

Zhou

Yao

Cui

Xue

Liang

. A content search method for security topics in microblog based on deep reinforcement learning. World Wide Web. 2020; 23(1): 75-101.

18.

Topal

Kurt

. Covert communication in cooperative NOMA networks. 2020 28th Signal Processing and Communications Applications Conference (SIU). IEEE. 2020. pp. 1-4.

19.

Shen

. A combined loss-based multiscale fully convolutional network for high-resolution remote sensing image change detection. IEEE Geoscience and Remote Sensing Letters. 2021; 19.

20.

Zhang

Kang

Zhang

Wen

Elhadad

Lei

. Speculation detection for Chinese clinical notes: impacts of word segmentation and embedding models. Journal of Biomedical Informatics. 2016; 60: 334-341.

21.

Pittner

Ehrenfellner

Monticelli

Zissler

Sänger

Stoiber

Steinbacher

. Postmortem muscle protein degradation in humans as a tool for PMI delimitation. International Journal of Legal Medicine. 2016; 130(6): 1547-1555.

22.

Aliyari

Rudzicz

Moghaddam

. Fast incremental LDA feature extraction. Pattern Recognition. 2015; 48(6): 1999-2012.

23.

Tvedebrink

Eriksen

Morling

. The multivariate dirichlet-multinomial distribution and its application in forensic genetics to adjust for subpopulation effects using the θ-correction. Theoretical Population Biology. 2015; 105: 24-32.

24.

Greff

Srivastava

Koutník

Steunebrink

Schmidhuber

. LSTM: A search space odyssey. IEEE Transactions on Neural Networks & Learning Systems. 2016; 28(10): 2222-2232.

25.

Zhang

Biswas

Adamchuk

. Implementation of a sigmoid depth function to describe change of soil pH with depth. Geoderma. 2017; 289: 1-10.

26.

Shin

Roth

Gao

Nogues

Yao

Mollura

Summers

. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning. IEEE Transactions on Medical Imaging. 2016; 35(5): 1285-1298.

27.

Yoshihiko

UEDA

Masajiro

ABE

Emi

ITO

. 910 fundamental dynamic experimental analysis on the tipping stability of crawler crane in jib luffing work: With forward lifted load on firm ground. Telecommunication Systems. 2015; 83(8): 1-12.

28.

Zhou

Pan

Tsang

. A deep learning framework for hybrid heterogeneous transfer learning. Artificial Intelligence. 2019; 275: 310-328.

29.

Mortensen

Langtangen

. High performance python for direct numerical simulations of turbulent flows. Computer Physics Communications. 2016; 203: 53-65.

30.

Zhang

Kang

Zhang

Wen

Elhadad

Lei

. Speculation detection for Chinese clinical notes: Impacts of word segmentation and embedding models. Journal of Biomedical Informatics. 2016; 60: 334-341.

31.

Bernard

Thiolliere

Saksida

Loukatou

Larsen

Johnson

Fibla

Dupoux

Daland

Cao

Cristia

. WordSeg: Standardizing unsupervised word form segmentation from text. Behavior Research Methods. 2020; 52(1): 264-278.

32.

White

Mattys

Stefansdottir

Jones

. Beating the bounds: Localized timing cues to word segmentation. Journal of the Acoustical Society of America. 2015; 138(2): 1214.

33.

Chen

. Classification of microblog users’ sentiments based on BERT-BILSTM-CBAM. CONVERTER. 2021; 281-287.

34.

Liu

Zeng

. Research on the emotional polarity of academic micro-blog knowledge transfer. Procedia Computer Science. 2021; 188: 86-92.

35.

Han

. Prediction and evolution of the influence of microblog entries in the context of terrorist events. Social Science Computer Review. 2021; 41(1).

36.

Fang

Chen

Lin

. A network based on deep interest network for Taobao click-through rate prediction. 2021 IEEE 2nd International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE). 2021. pp. 978-981.

User characteristics	Number of fans; degree of attention; number of microblogs; whether or not there is certification;
	level; number of microblogs transferred
Microblog features	Are there any pictures or videos; number of reposts; number of comments; number of likes

The microblog sentiment analysis based on latent dirichlet allocation and deep learning approaches

Abstract

Keywords

1. Introduction

2. Literature review

2.1 Deep learning status

2.2 The adoption status of neural network in sentiment analysis

3. Materials and methods

3.1 Sentiment analysis of reposting microblog under LDA model

3.2.1 LSTM sentiment classification model

4. Results

4.1 Experimental data sets

4.6 Sentiment analysis results of user reposting microblog

Table 5 Comparison of the results of sentiment binary classification between CNN model and LSTM model

5. Conclusion

5.1 Research summary

5.2 Limitations of this research

Footnotes

Funding

References

Table 5
Comparison of the results of sentiment binary classification between CNN model and LSTM model