Research on Red Cultural Inheritance and Application of SVM Support Vector Machine in Sentiment Analysis

Latent Semantic Analysis (LSA) is also known as Latent Semantic Indexing [27]. The essence of the latent semantic analysis method is to find out the real semantics of the words in the document, and then to mine out the document’s word-independent topics, i.e., latent semantic topics. Thus, it solves the problems and deficiencies associated with the inability to consider latent semantics in the vector space model. Specifically, a large collection of documents is modeled using a reasonable dimension. And the process of representing the documents in this space is based on Singular Value Decomposition (SVD) and dimensionality reduction. Dimension reduction is the most important step in LSA analysis. Through dimensionality reduction, the “noise”, i.e., irrelevant information in the document is removed, so that the semantic structure is gradually presented. Compared to the VSM model, dimensionality reduction reduces computational effort and clears the semantic relationship.

With SVD, the previous document-word covariance matrix A can be subset into three matrices: (6)X=U∑VT$$X = U\sum {{V^T}}$$

where U and V are the west matrices of X and Σ is the singular value matrix.

For matrix Σ, the lower order approximation matrix of matrix X is obtained by setting its smallest r − k singular values to 0, which gives Σ_k Then the lower order approximation matrix of matrix 8 is obtained by Eq. (7): (7)Ck=U∑kVT$${C_k} = U\sum\limits_k {{V^T}}$$

The general steps of LSA can be summarized as follows:

Step 1: Analyze the document collection and establish the “document-feature word” matrix (TD).

Step 2: Perform singular value diversity (SVD) on the TD matrix.

Step 3: Perform dimensionality reduction on the matrix after SVD singular value diversity, that is, the low-order approximation mentioned earlier.

Step 4: Use the reduced matrix to construct the latent semantic space or reconstruct the TD matrix.

Categorizing text is also a process of creating a mapping of text to categories. It maps the text to be categorized into the existing categories, and the mapping can be a one-to-one mapping or a one-to-many mapping. Because a text can relate to more than one topic. The mathematical description is as follows: (8)f:A→B$$f:A \to B$$

Where A is the collection of text data to be classified, B is the collection of categories in the classification system, and f is the mapping rule.

The mapping rule f of automatic text classification is the discriminative formula and discriminative rule established by the system according to the data information of a number of samples in each category that has been mastered to summarize the regularity of classification. Then when a new text is encountered, the category to which the text belongs is determined according to the summarized discrimination rules. 1)

Text Preprocessing

The primary work of text preprocessing is to deal with noise information and irregular information. Take web page text as an example, there will be a large amount of HTML markup information after it is acquired. This information is used for the layout and display of the web page text, but it is basically of no value to the text. After removing this kind of meaningless interference information, it is the second step in text preprocessing - word separation. The difficulty of word separation is to determine the smallest element of the vocabulary, that is, the most basic semantic unit.

Therefore, by converting all capital letters appearing in the text into lowercase forms, and by using non-alphabetic characters such as spaces, punctuation marks, etc. as separators, the text can be easily converted into a tabular list composed of semantic units (words).

where f_ij denotes the frequency of occurrence of word i in document j, N is the total number of texts in the text collection, and n_i is the total number of occurrences of word i in the text collection. 3)

Text features

When representing text as a feature vector, the original feature set consists of all words that appear in the text set. There are two main purposes for performing feature dimensionality reduction. First, if training and classification are performed directly on such a high-dimensional feature space, the amount of computation is too large. Dimensionality reduction can improve the execution efficiency and running speed of the program. Second, all words have different meanings for text categorization. Some generalized lexical entries, which are prevalent in all classes, contribute little to the classification. Phrases that appear in a specific class with a large proportion and in other classes with a small proportion contribute to text categorization, and dimensionality reduction improves the generalization ability of the classifier.

Feature dimensionality reduction is the selection of a true subset from the original document set T={t1,t2,⋯ts}$$T = \left\{ {{t_1},{t_2}, \cdots {t_s}} \right\}$$: (10)T′={tn,tp2,⋯tps}$$T\prime = \left\{ {{t_n},{t_{{p_2}}}, \cdots {t_{{p_s}}}} \right\}$$

satisfies p_s ≤ s. where s is the size of the original feature set and p_s is the size of the feature set after dimensionality reduction. The criterion for selection is that the classification accuracy can be effectively improved after feature dimensionality reduction. Feature dimensionality reduction does not change the nature of the original feature space, but only selects a part of the important features from the original feature space to form a new low-dimensional space.

In the all-media era, certain technical conditions and changes in the social environment have raised new problems and brought new challenges to the inheritance of red culture. To do a good job of red cultural heritage work, we should summarize the past results on the basis of an in-depth study of the current outstanding problems, analyze the new trends, the new environment contains opportunities for development, and then grasp the favorable factors to improve the relevant work.

In recent years, the form of red culture carriers has been further expanded, gradually developing into various thematic educational activities that incorporate the characteristics of the new era. For example, in 2018, the Central Committee of the Communist Youth League organized and launched a series of learning activities for the youth, which included online knowledge contests, essay contests, speech contests, fun question and answer contests, and other forms, to further promote the normalization of red culture education, which is rich in connotations and has a remarkable effect.

Through TV programs, graphic news, webcasts and other channels, red culture has entered grass-roots organizations in various fields such as enterprises and public institutions, rural areas and communities, gradually taking the point to lead the way, breaking the audience limitations, and creating a trend in the society, so that more people come into contact with and enjoy red culture.

In recent years, the Party and the State have attached great importance to the inheritance of red culture and given support in various aspects, but deficiencies in theoretical construction and social synergy still exist. People’s understanding of red culture is a little mixed, and their attitudes towards inheritance work are also very different. Analyzing the reasons, the inheritance of red culture mainly exists in the following four aspects. 1)

Changes in audience thinking

From the audience’s point of view, the red culture inheritance work is facing the deconstruction of the negative factors in the modern ideological concepts. Since the reform and opening up, politics, economy, culture, society and other fields have ushered in great changes, and the ever-changing information technology and high-tech power has profoundly changed all aspects of people’s food, clothing, housing and transportation. To a certain extent, this has reshaped people’s values, ways of thinking, and modes of action, and has brought new challenges to the inheritance of red culture.

A social media webpage is used as the experimental data and analysis object, and the experiments use Python’s BeautifulSoup as the page element crawling and parsing tool, and adopt the parallelization crawling strategy to obtain the Chinese text data.

The text of webpages within a certain period of time is collected as the original dataset. Due to the wide range of web page text data types, the filtering measures in the data crawling stage are not good enough to obtain the experimental data. Therefore, the crawled data needs to be manually filtered and labeled.

The experimental data contains the following information: user ID, homepage address, text content, release date (specific time period), and URL. It is important to state that the methods and data used in the text are privacy-protected.

Although the number of webpage texts is certain, the complexity and diversity of Chinese expressions lead to a certain degree of redundancy in the collected data. Therefore, the data needs to be preprocessed so as to extract sentiment information from unstructured data.

In order to more accurately detect whether the text content of web pages contains red cultural inheritance tendency, this paper focuses on the field of psychology, while taking into account the Internet terminology and Internet spoken expressions, etc., collects words and other expressions related to red cultural inheritance, and establishes a hierarchical classification scheme that includes different mental states.

Based on the sentiment classification scheme, each webpage text is manually labeled according to its content.

After labeling each data text for classification and initial preprocessing, the processed data is vectorized using Word2Vec, i.e., each preprocessed web page text is converted to Vi(x1,x2,x3,x4…xk)$$Vi\left( {{x_1},{x_2},{x_3},{x_4} \ldots {x_k}} \right)$$ and computed using a matrix. In this experiment, similarity between words is also taken into account to find similar words in different webpage texts. As for the calculation of similarity, this paper adopts the Euclidean distance calculation method, as in Equation (11): (11)S=∑i=1k(x1−x2)2$$S = \sum\limits_{i = 1}^k {\sqrt {{{\left( {{x_1} - {x_2}} \right)}^2}} }$$

where similarity is p = 1/(1 + S). For the number of similar words that are assigned the same weight and classified in the same subcategory, i.e., Eq. k value, can be customized.

The basic model of SVM is to find the best hyperplane in a particular space for solving binary classification problems. However, with the deepening of research, the model can be used to solve nonlinear problems after the introduction of kernel function. In this paper, a 3-layer hierarchical classification model is constructed based on SVM according to the sentiment classification scheme to predict the probability of web users’ tendency to inherit red culture. The model uses the default RBF kernel function. In this case, the classification target corresponding to each layer of classification is based on the neighboring layer of classification. For the first layer of classification the segmentation hyperplane representation is shown in Equation (12): (12)f1(x)=ω1Tφ(x)+b1$${f_1}(x) = \omega _1^T\varphi (x) + {b_1}$$

where ω_t is the normal vector, which determines the direction of the separating plane. b_t is the displacement term, which determines the distance between the separating plane and the origin. The goal of this separating hyperplane classification is to determine whether the web page text content is emotionally relevant. Similarly, the separation hyperplane representation for the second level of classification is shown in Equation (13): (13)f2(x)=ω2Tφ(x)+b2$${f_2}(x) = \omega _2^T\varphi (x) + {b_2}$$

Where the constant b₂ has the same meaning as the corresponding meaning of equation (12), the determination of this hyperplane is based on the first level of classification. The main goal is used to classify web texts containing negative and non-negative emotions. The third layer of classification is the most important part of this experiment. By filtering the first two layers of classification, the data obtained is for each layer of classification, which can be transformed into its dyadic problem solving in order to get the hyperplane of maximum interval division. In this experiment, the Lagrange multiplier algorithm is utilized to obtain its dual problem, which is shown in Equation (14): (14)L(ω,b,α)=12‖ω‖2+∑αi(1−yi(ωTxi+b))$$L(\omega ,b,\alpha ) = \frac{1}{2}{\left\| \omega \right\|^2} + \sum {{\alpha _i}} \left( {1 - {y_i}\left( {{\omega ^T}{x_i} + b} \right)} \right)$$

The heart of the method lies in the derivation of the function L with respect to the relevant variables, and the dyadic problem that is ultimately transformed into is Eq. (15): (15)maxα∑i=1mαi−12∑i=1m∑i=0mαiαjyiyjxiTxj s.t.∑i=0mαiyi,i=1,2,3⋯m$$\begin{array}{l} {\max _\alpha }\sum\limits_{i = 1}^m {{\alpha _i}} - \frac{1}{2}\sum\limits_{i = 1}^m {\sum\limits_{i = 0}^m {{\alpha _i}} } {\alpha _j}{y_i}{y_j}x_i^T{x_j} \\ s.t.\sum\limits_{i = 0}^m {{\alpha _i}} {y_i},i = 1,2,3 \cdots m \\ \end{array}$$

After several experiments, it is found that parameter optimization can have a significant impact on the data analysis results. For SVM models with RBF kernel function, parameters C and γ need to be optimized. Among them, C is the tolerance level of classification error, and γ is a self-contained parameter of the kernel function. The higher value of parameter C indicates the more stringent requirements for model accuracy, and parameter γ determines the distribution of values after the original data is mapped into the feature space of different dimensions. With the use of the kernel function, the dyadic problem corresponding to Eq. (15) can be converted to Eq. (16): (16)maxα∑i=1mαi−12∑i=1m∑i=0mαiαjyiyjφ(xiT)φ(xj) s.t.∑i=0mαiyi,i=1,2,3⋯m$$\begin{array}{c} {\max _\alpha }\sum\limits_{i = 1}^m {{\alpha _i}} - \frac{1}{2}\sum\limits_{i = 1}^m {\sum\limits_{i = 0}^m {{\alpha _i}} } {\alpha _j}{y_i}{y_j}\varphi \left( {x_i^T} \right)\varphi \left( {{x_j}} \right) \\ s.t.\sum\limits_{i = 0}^m {{\alpha _i}} {y_i},i = 1,2,3 \cdots m \\ \end{array}$$

where φ(x) is the kernel function used.

A webpage comment text dataset is used and 1500 training data are selected first. Positive and negative text data are balanced and used to train the support vector machine. Radial basis kernel function is used for the kernel parameters of the support vector machine. One set is using the parameters obtained by default in the model of the experimental tool, while the other set is using the optimal combination of parameters found by the genetic algorithm. Three test sets are created, each with 200 test data with identical categories. The two sets of classifier models for sentiment categorization are evaluated for their effectiveness.

Here is a comparison of the performance of the optimized SVM classifier. There are three sets of test data, and the comparison of the two classification results is shown in Figure 3.

Figure 3.

Comparison of the classification results

The analysis on the experimental data shows that in the three evaluation metrics of Accuracy, Recall and F-measure value, the Hierarchical Support Vector Machine classification results are better and there is some performance improvement in all the three metrics. The mean values of the three evaluation metrics of the hierarchical support vector machine are around 90, which is higher than the mean value of the SVM model before optimization. And the results of all three sets of data are better than the support vector machine model with default parameters, indicating that the RBF kernel function can find a better combination of parameters in the results of parameter optimization search, which can effectively improve the classification performance of the classifier.

1)

NLPIR Chinese Word Segmentation System

In this paper, NLPIR participle system is used. Among them, lexical annotation, named entity recognition, and user dictionary are all within the functional scope of Chinese participle system. It supports GBK encoding, UTF8 encoding, and BIG5 encoding.

Realization of automatic adaptation of participles requires automatic discovery of new words based on information cross entropy combined with feature phrases in slightly long text sentences, and the distribution model is realized by automatic adaptation to test the predicted linguistic probability, which is the function of emotional new word discovery and automatic adaptation of participles.

STEP2: The new words recognized by the segmentation tool need to go to the dictionary to find the polarity and intensity.

STEP3: The construction of feature vectors need to transform the intensity and polarity, stipulating that the intensity are divided by ten to get the value of sentiment word intensity. For polarity greater than 1 is designated as -1 (i.e., negative), polarity equal to 0 is designated as 0 (i.e., neutral), and polarity equal to 1 is designated as 1 (i.e., positive).

STEP4: The SVM classifier is used to train the data, and the vectorized training data needs to be trained to construct the viewpoint sentence extraction model. The constructed model is then used to classify the test corpus.

The crawled daily comment statements are brought into the comment1 of the hierarchical support vector machine for initial evaluation. If the output result of comment1 is 1, the output is changed to comment2. If the output result is 0, go to comment3. If the output result of comment2 is 1, it indicates a positive comment. If the result is 0, go to comment4. If the output result of comment3 is 1, it indicates a negative comment. If the result is 0, it indicates a negative review. If the output of comment4 is 1, it indicates a positive comment. If the result is 0, it means a neutral comment. Table 2 lists the detailed results of information collection. During the survey period from May 1 to May 9, the total crawl data of positive comments, neutral comments, and negative reviews were 4360, 1462, and 2953, respectively. The proportion of positive comments may be due to the fact that the survey period coincides with Labor Day and Youth Day, and online comments show a positive communication atmosphere. The daily accuracy of the hierarchical support vector machine proposed in this paper for daily comment statements is in the range of 85% to 89%, respectively, and the accuracy of daily comment classification is high.

An overall count of each sentiment type in the comment data reveals that positive sentiments are more prominent, mainly due to the positive guidance of traditional festivals and the promotion of red cultural festivals in the official media. The overall sentiment tends to be positive. The trend of the number of posted sentiments of daily comments is shown in Figure 4. From the plumb line graph of the number of daily comments on sentiment, it can be seen that May 1-May 4 is dominated by positive sentiment comments, and the discussion about Labor Day and Youth Day dominates this social media network.

Figure 4.

Daily comments on the number of emotional trends

In order to be able to further understand the main concerns of web users about the comments on the traditional festival of red culture - May 4 Youth Day, this paper counts the top 20 high-frequency keywords appearing in the comments. The high-frequency words of online comments are shown in Table 3.

The top 10 keywords are “Youth Propaganda”, “Holiday”, “Youth”, “Youth Day Message”, “May Fourth Commemorative Activities”, “May Fourth Movement”, “Origin of Youth Day”, “Youth Day Greetings”, “Youth Day Activities”, and “The Historical Significance of the May Fourth Movement”. These high-frequency words can reflect that most netizens show a high degree of attention to issues such as red cultural traditional festivals, and hope to participate in the construction of red cultural traditional festivals and inherit and pass on red culture.

May Fourth Youth Day is not only a commemorative day, but also a cultural symbol with symbolic significance. In different historical periods, the red cultural connotations covered by May Fourth Youth Day have guided the development direction of the youth movement. May 4 Youth Day is one of the symbols of the contemporary inheritance and development of red culture.

Although “Internet + red culture” has development potential, at present, the social network platform’s preference for the dissemination of entertainment information tends to crowd out the space for the dissemination of red culture. At the same time, the fragmented and fast-food narrative logic of social network platforms will dismember the wholeness and profoundness of red culture.

In order to better utilize the Internet to spread red culture and realize the “Internet inheritance of red culture”, policymakers need to fully explore its economic value and promote the high-quality development of red culture industry.

Accelerate specialized legislation and judicial practice for the protection of red historical and cultural resources. There is no relevant basis for directly pursuing legal responsibility for malicious rumor-mongering and smearing of red culture on the Internet.

(c) Implementing regularized special operations to maintain a clear order in cyberspace. In response to harmful content such as “distortions of red history, denigration of the Party’s guidelines and policies, and achievements in social construction”, the public security authorities and the competent departments in charge of Internet and information technology have set up a special operation to combat false rumours about red culture and history. A mechanism for reviewing and guiding red cultural information has been set up, with increased supervision of platforms and dissemination channels, and a special area for reporting false information on red history and culture has been established.

Promoting the development of red culture in cyberspace through “Internet+”, enhancing the digital influence of red culture, educating and guiding netizens to follow socialist core values, and creating a new way of educating people through red culture on the Internet. Through “Internet+”, we have strengthened the placement of public service advertisements in cities, and made use of urban public broadcasting systems, such as bus mobile radio, television, electronic bulletin boards, giant-screen advertisements, and theaters to broadcast red stories and commemorative videos. Red cultural education is socialized through the network, providing citizens with extensive and innovative ideological and political education, and building a good urban environment for red education. Meanwhile, mainstream media actively organize cross-media linkage and hold widely attended programs to promote knowledge of red culture. Through central media TV placement, linked to cell phone question and answer prizes or promotion mechanisms, to promote the participation of the whole society, create viewing hot spots, and create a red topic.