An In-depth Analysis of Data Science Methods on the Path of Women’s Consciousness Awakening under the Cultural System of Marxist Chineseization

In the CBOW model, word ω_t is unknown, but the context of ω_t ω_t–2, ω_t–1, ω_t+1, ω_t+2 is known, the core idea of CBOW is to use the context of ω_t to predict the unknown ω_t, the model through the training progress, constantly update the parameters, use the gradient ascent method to maximize the objective function, output the most likely ω_t and the updated value of the word vector, and sum the updated word vector and the initial vector to obtain the final trained word vector of the CBOW model [19].

The cbow model structure consists of a three-layer network. The input layer inputs the initial word vectors with known contexts, the middle projection layer sums up the vectors input from the input layer, and in the last layer the output layer model modifies the values of each parameter in the model and the vector values initially input to the model by solving the maximum value of the objective function, inversely modifying the values of each parameter in the model and the initial input to the model, and when the objective function is maximized the current word ω can be predicted, and at the same time the trained word vectors can also be obtained. Using Context(ω) to predict the current word ω, the objective function is defined as follows: (1) L=∑ω∈Clogp(ω|Context(ω)) \[L=\sum\limits_{\omega \in C}{\log p\left( \omega \left| Context\left( \omega \right) \right. \right)}\] Where: ω denotes any word in corpus C, and Context(ω) denotes the context of ω.

The cbow model based word2vec algorithm trains word vectors as follows:

Input layer: a random vector of size m of the context Context(ω) of input ω is updated as training progresses.

Projection layer: the initial word vectors of context Context(ω) input to the input layer are directly summed as input to the hidden layer, i.e., Xω=∑i=12cV(Context(ω)i)∈Rm${{X}_{\omega }}=\sum\limits_{i=1}^{2c}{V}\left( Context{{\left( \omega \right)}_{i}} \right)\in {{R}^{m}}$.

Output layer: In order to simplify the complex multiclassification problem, the output layer of the model is designed as a Huffman tree structure, which can transform the multiclassification problem into a biclassification problem after multiple bisections. Figure 1 shows the output layer structure of the cbow model, bisection is performed once at each non-leaf node in the Huffman tree, to the left or to the right, and the corresponding output value is represented by 0 or 1. The leaf nodes of the Huffman tree are all the words in the corpus, and the weights corresponding to the leaf nodes are the number of times the word corresponding to the node appears in the corpus. The specific process of solving the objective function based on the cbow model optimized by the hierarchl softmax algorithm is to search from the root node of the Huffman tree all the way to the corresponding leaf node where the word is located, and multiply all the probabilities on the whole path of the search.

Figure 1.

Sample cbow model

In the figure, the context of female thoughts is used to predict the probability of “female thoughts”, and when the objective function is maximized, the model searches from the leaf node that maximizes the objective function to the root node while modifying the parameter values and the initial word vector.

The probability of going left or right in the Huffman tree is calculated according to equations (2) and (3). The probability of positive instances to the right is calculated according to equation (2) and the probability of negative instances to the left is calculated according to equation (3): (2) σ(XωTθ)=11+e−XωTθ \[\sigma \left( X_{\omega }^{T}\theta \right)=\frac{1}{1+{{e}^{-X_{\omega }^{T}\theta }}}\] (3) 1−σ(XωTθ)=e−XωTθ1+e−XωTθ \[1-\sigma \left( X_{\omega }^{T}\theta \right)=\frac{{{e}^{-X_{\omega }^{T}\theta }}}{1+{{e}^{-X_{\omega }^{T}\theta }}}\] (4) L=∑ω∈Cp(ω|Context(ω))=∏j=2lωp(djω|Xω,θj−1ω) \[L=\sum\limits_{\omega \in C}{p}\left( \omega \left| Context\left( \omega \right) \right. \right)=\prod\limits_{j=2}^{{{l}^{\omega }}}{p}\left( d_{j}^{\omega }\left| {{X}_{\omega }} \right.,\theta _{j-1}^{\omega } \right)\] (5) L=∑ω∈C∑j=2lω{ (1−djω)⋅log[ σ(Xωτθj−1ω) ]+djω⋅log[ 1−σ(XωTθj−1ω) ] } \[L=\sum\limits_{\omega \in C}{\sum\limits_{j=2}^{{{l}^{\omega }}}{\left\{ \left( 1-d_{j}^{\omega } \right)\cdot \log \left[ \sigma \left( X_{\omega }^{\tau }\theta _{j-1}^{\omega } \right) \right]+d_{j}^{\omega }\cdot \log \left[ 1-\sigma \left( X_{\omega }^{T}\theta _{j-1}^{\omega } \right) \right] \right\}}}\]

Use the gradient ascent method to obtain a partial derivative of the above equation with respect to θj−1ω$\theta _{j-1}^{\omega }$: (6) ∂L(ω,j)∂θj−1ω=∂∂θj−1ω{ (1−djω)⋅log[ σ(Xωτθj−1ω) ]+djω ⋅log[ 1−σ(Xωτθj−1ω) ] } \[\begin{align} & \frac{\partial L\left( \omega ,j \right)}{\partial \theta _{j-1}^{\omega }}=\frac{\partial }{\partial \theta _{j-1}^{\omega }}\left\{ \left( 1-d_{j}^{\omega } \right)\cdot \log \left[ \sigma \left( X_{\omega }^{\tau }\theta _{j-1}^{\omega } \right) \right]+d_{j}^{\omega } \right. \\ & \left. \cdot \log \left[ 1-\sigma \left( X_{\omega }^{\tau }\theta _{j-1}^{\omega } \right) \right] \right\} \end{align}\]

The updated expression for θj−1e$\theta _{j-1}^{e}$ is: (7) θj−1ω=θj−1ω+η[ 1−djω−σ(XωTθj−1ω) ]Xω \[\theta _{j-1}^{\omega }=\theta _{j-1}^{\omega }+\eta \left[ 1-d_{j}^{\omega }-\sigma \left( X_{\omega }^{T}\theta _{j-1}^{\omega } \right) \right]{{X}_{\omega }}\]

Similarly for Eq. (5) with respect to X_ω for partial derivatives: (8) ∂L(ω,j)∂Xω=[ 1−djω−σ(XωTθj−1ω) ]⋅θj−1ω \[\frac{\partial L\left( \omega ,j \right)}{\partial {{X}_{\omega }}}=\left[ 1-d_{j}^{\omega }-\sigma \left( X_{\omega }^{T}\theta _{j-1}^{\omega } \right) \right]\cdot \theta _{j-1}^{\omega }\]

The updated expression for V_w is: (9) Vω:=Vω+η∑j=2lω([ 1−djω−σ(XωTθj−1ω) ]θj−1ω) \[{{V}_{\omega }}:={{V}_{\omega }}+\eta \sum\limits_{j=2}^{{{l}^{\omega }}}{\left( \left[ 1-d_{j}^{\omega }-\sigma \left( X_{\omega }^{T}\theta _{j-1}^{\omega } \right) \right]\theta _{j-1}^{\omega } \right)}\]

The idea of skip-gram model is more similar to the idea of cbow model. skip-gram model is that the input word ω_t is known, while the context of the word is unknown, the model uses the current word to predict the probability of its context ω_t–2, ω_t–1, ω_t+1, ω_t+2 appearing. skip-gram model also consists of a three-layer network structure of the input layer, the projection layer, and the output layer [20].

Input layer: the input intermediate word ω corresponds to a dimension of m random initialization vector v(ω) ∈ R^m.

Projection layer: in fact, this layer does not have much significance, mainly to correspond to the cbow model structure.

Output layer: also corresponds to a Huffman tree can be analogous to Figure 1, through the Huffman tree can be transformed into a complex multi-classification problem into binary classification. The skip-gram model based on hierarchl softmax optimization is to transform the probability of ω contexts after softmax normalization into a Huffman tree, where the leaf nodes correspond to the words in the corpus lexicon, and the non-leaf nodes decide to which child node the word vectors are assigned. So the objective function of the skip-gram model based on hierarchl softmax optimization is computed as the product of the probabilities of the non-leaf nodes on the paths traveled from the root node of the Huffman tree to the leaf nodes. The objective function of this model is defined as follows: (10) L=∑ω∈Clogp(Context(ω)|ω) \[L=\sum\limits_{\omega \in C}{\log }p\left( Context\left( \omega \right)\left| \omega \right. \right)\] (11) L=∑ω∈Cp(Context(ω)|ω)=∏j=2tup(dju|V(ω),θj−1u) $L=\sum\limits_{\omega \in C}{p}\left( Context\left( \omega \right)\left| \omega \right. \right)=\prod\limits_{j=2}^{{{t}^{u}}}{p}\left( d_{j}^{u}\left| V(\omega ) \right.,\theta _{j-1}^{u} \right)$ (12) p(dju|V(ω),θj−1u)=[ σ(V(ω)θj−1u) ]1−djω⋅[ 1−σ(V(ω)θj−1u) ]1−djω $p\left( d_{j}^{u}\left| V(\omega ) \right.,\theta _{j-1}^{u} \right)={{\left[ \sigma \left( V(\omega )\theta _{j-1}^{u} \right) \right]}^{1-d_{j}^{\omega }}}\cdot {{\left[ 1-\sigma \left( V(\omega )\theta _{j-1}^{u} \right) \right]}^{1-d_{j}^{\omega }}}$ (13) p(dju|V(ω),θj−1u)=[ σ(V(ω)θj−1u) ]1−djω⋅[ 1−σ(V(ω)θj−1u) ]1−djω \[p\left( d_{j}^{u}\left| V\left( \omega \right) \right.,\theta _{j-1}^{u} \right)={{\left[ \sigma \left( V\left( \omega \right)\theta _{j-1}^{u} \right) \right]}^{1-d_{j}^{\omega }}}\cdot {{\left[ 1-\sigma \left( V\left( \omega \right)\theta _{j-1}^{u} \right) \right]}^{1-d_{j}^{\omega }}}\] (14) L=∑ω∈C∑u∈Context(ω)∑j=2tω{ (1−djω)⋅log[ σ(V(ω)Tθj−1u) ] +djω⋅log[ 1−σ(V(ω)Tθj−1u) ] } $\begin{align} & L=\sum\limits_{\omega \in C}{\sum\limits_{u\in Context\left( \omega \right)}{\sum\limits_{j=2}^{{{t}^{\omega }}}{\left\{ \left( 1-d_{j}^{\omega } \right)\cdot \log \left[ \sigma \left( V{{\left( \omega \right)}^{T}}\theta _{j-1}^{u} \right) \right] \right.}}} \\ & \left. +d_{j}^{\omega }\cdot \log \left[ 1-\sigma \left( V{{\left( \omega \right)}^{T}}\theta _{j-1}^{u} \right) \right] \right\} \end{align}$

The update formula for the parameter vector θj−1u$\theta _{j-1}^{u}$ obtained by partial derivation of equation (14) with respect to θj−1u$\theta _{j-1}^{u}$ using the gradient ascent method is: (15) θj−1∞=θj−1∞+η[ 1−djω−σ(V(ω)Tθj−1u) ]⋅V(ω) \[\theta _{j-1}^{\infty }=\theta _{j-1}^{\infty }+\eta \left[ 1-d_{j}^{\omega }-\sigma \left( V{{\left( \omega \right)}^{T}}\theta _{j-1}^{u} \right) \right]\cdot V\left( \omega \right)\]

The update formula for the parameter vector V(ω) obtained by partial derivation of equation (14) with respect to V(ω) using the gradient ascent method is: (16) V(ω)=V(ω)+η∑u∈Contect(ω)∑j=2tω([ 1−djω−σ(V(ω)Tθj−1u) ]⋅θj−1u) $V\left( \omega \right)=V\left( \omega \right)+\eta \sum\limits_{u\in Contect\left( \omega \right)}{\sum\limits_{j=2}^{{{t}^{\omega }}}{\left( \left[ 1-d_{j}^{\omega }-\sigma \left( V{{\left( \omega \right)}^{T}}\theta _{j-1}^{u} \right) \right]\cdot \theta _{j-1}^{u} \right)}}$

In this section, based on the Word2Vec tool and the LDA topic model, this paper will introduce a distance metric, EMD distance formula, which is widely used in the image processing field but less used in the text mining field, to build a model of W2v_dist algorithm. First, the feasibility of using the EMD formula to measure text distance is evaluated.

EMD distance is also known as land movement distance, and EMD distance formula is commonly used to solve the optimal solution of transportation problems and calculate the similarity of images. It is widely used in the fields of image processing and computer vision.

The EMD distance provides a good quantification of the minimum cost required to transform histogram P into histogram Q, i.e., the degree of similarity between the two. Assumptions P = {p₁,p₂,…p_n}, Q = {q₁,q₂,…q_n}. The comparison leads to the ground distance matrix D=[ di,j ]$D=\left[ {{d}_{i,j}} \right]$. where d_i,j denotes the distance traveled from the ith data bucket in P to the jth data bucket in Q. Let f_i,j ∈ F be the size of the probabilistic flow from p_i to q_j, i.e., f_[i,j] = |p_i – q_j|. EMD distance essentially solves a handling problem, i.e., the total amount remains the same before and after transportation. Therefore, it is common practice to normalize P and Q so that their totals are 1. This leads to Equation (20): (20) EMD(P,Q)=min(∑i=1n∑j=1nfijdij)s.t.∀i:∑jfij=pi∀j:∑ifij=qj∀i,j:fij≥0 \[\begin{align} & EMD\left( P,Q \right)=\min \left( \sum\limits_{i=1}^{n}{\sum\limits_{j=1}^{n}{{{f}_{ij}}}}{{d}_{ij}} \right) \\ & s.t.\forall i:\sum\limits_{j}{{{f}_{ij}}}={{p}_{i}} \\ & \forall j:\sum\limits_{i}{{{f}_{ij}}}={{q}_{j}} \\ & \forall i,j:{{f}_{ij}}\ge 0 \\ \end{align}\]

From equation (20), EMD exhibits the least consumable cost of transforming histogram P into histogram Q.

EMD distance utilizes the features of the images to represent the distance of the images. Suppose there exist two images P and Q, p_i and q_j denote the features of these two images respectively. n_i denotes the weight values of the features. Define the distance matrix [ dij ]$\left[ {{d}_{ij}} \right]$ for feature P and feature Q with Size as 1*m. Suppose there exists matrix F=[ fij ]$F=\left[ {{f}_{ij}} \right]$, f_ij representing the amount of transfer from p_i to q_j. Then there exists minimization cost function: (21) Work(P,Q,F)=∑i=1l∑j=1mdijfij \[Work\left( P,Q,F \right)=\sum\limits_{i=1}^{l}{\sum\limits_{j=1}^{m}{{{d}_{ij}}}}{{f}_{ij}}\]

A normalized expression of Eq. (21) yields the EMD distance equation: (22) EMD(P,Q)=∑i=1l∑j=1mdijfij∑i=1l∑j=1mfij \[EMD\left( P,Q \right)=\frac{\sum\limits_{i=1}^{l}{\sum\limits_{j=1}^{m}{{{d}_{ij}}}}{{f}_{ij}}}{\sum\limits_{i=1}^{l}{\sum\limits_{j=1}^{m}{{{f}_{ij}}}}}\]

EMD distance can also be expressed in terms of the distance between texts in terms of the features in the text. From the analysis above, it can be seen that the topic is the feature of the text, while the words are the features of the topic. In this paper, we will use word vector and EMD distance to calculate the distance between topics. Then, the distance between topics combined with EMD formula is used to get the distance between texts.

The first step in the implementation of the W2v_dist algorithm is to train the word vectors. Training Word2Vec word vectors means that the researcher utilizes the Skip-gram model or the CBOW model in the Word2Vec tool to train a corpus in a particular domain. The ultimate goal is to obtain word vectors for each word in that corpus.

This paper describes the principle of LDA is that each topic can in turn be viewed as a multinomial distribution of a number of words. That is, words are features of topics. Then, each topic vector can be represented in the following form: (23) vec(Top)=(weight(w1),weight(w2),weight(w3),…,weight(wn)) \[vec\left( Top \right)=\left( weight\left( {{w}_{1}} \right),weight\left( {{w}_{2}} \right),weight\left( {{w}_{3}} \right),\ldots ,weight\left( {{w}_{n}} \right) \right)\] Where weight(w_i) is the weight value occupied by the ind feature word. If common distance calculation formulas such as Euclidean distance and Ma distance are used to calculate the distance between topics, the correlation between words is ignored.

In order to better utilize the semantic information to calculate the distance between topics, this paper proposes the following formula in combination with EMD distance: (24) consT(Top1,Top2)=min∑i=1n∑j=1ntransW(i,j)⋅consW(Wordi,Wordj)s.t.transW(i,j)≥0;∑i=1ntransW(i,j)=WWQj;∑i=1ntransW(i,j)=WWPi∑i=1n∑j=1ntransW(i,j)=∑i=1NWWPi=∑i=1NWWQi=1i,j=1,2,…n \[\begin{align} & consT\left( To{{p}_{1}},To{{p}_{2}} \right)=\min \sum\limits_{i=1}^{n}{\sum\limits_{j=1}^{n}{transW\left( i,j \right)}}\cdot consW\left( Wor{{d}_{i}},Wor{{d}_{j}} \right) \\ & s.t. \\ & transW\left( i,j \right)\ge 0; \\ & \sum\limits_{i=1}^{n}{trans}W\left( i,j \right)=WW{{Q}_{j}}; \\ & \sum\limits_{i=1}^{n}{trans}W\left( i,j \right)=WW{{P}_{i}} \\ & \sum\limits_{i=1}^{n}{\sum\limits_{j=1}^{n}{trans}}W\left( i,j \right)=\sum\limits_{i=1}^{N}{W}W{{P}_{i}}=\sum\limits_{i=1}^{N}{W}W{{Q}_{i}}=1 \\ & i,j=1,2,\ldots n \\ \end{align}\]

In Eq. (24), Top₁ = (WWP₁,WWP₂,…,WWP_n), Top₂ = (WWQ₁,WWQ₂,…,WWQ_n). WWP_i and WWQ_i denote the weight values of the same feature word in two topics. The offset of moving from WWP_i to WWQ_j can be represented by transW(i,j). Normalization is performed for transW(i,j).

In the previous section, this paper pointed out that themes can be viewed as features of a text. Each text can be viewed as a vector of low dimensions, and each dimension of the vector represents a topic. The dimension value is the weight of the topic in the text. Therefore, this paper again applies the EMD distance formula and defines the text distance metric formula as follows: (25) W2−dist(Doc1,Doc2)=min∑i=1N∑j=1NtransT(i,j)⋅consT(Topi,Topj)s.t.transT(i,j)≥0∑i=1NtransT(i,j)=TWQj∑i=1NtransT(i,j)=TWPi∑i=1N∑j=1NtransT(i,j)=∑i=1NTWPi=∑i=1NTWQi=1i,j=1,2,…N \[\begin{align} & W{{2}_{-}}dist\left( Doc1,Doc2 \right)=\min \sum\limits_{i=1}^{N}{\sum\limits_{j=1}^{N}{trans}}T\left( i,j \right)\cdot consT\left( To{{p}_{i}},To{{p}_{j}} \right) \\ & s.t. \\ & transT\left( i,j \right)\ge 0 \\ & \sum\limits_{i=1}^{N}{trans}T\left( i,j \right)=TW{{Q}_{j}} \\ & \sum\limits_{i=1}^{N}{trans}T\left( i,j \right)=TW{{P}_{i}} \\ & \sum\limits_{i=1}^{N}{\sum\limits_{j=1}^{N}{trans}}T\left( i,j \right)=\sum\limits_{i=1}^{N}{T}W{{P}_{i}}=\sum\limits_{i=1}^{N}{T}W{{Q}_{i}}=1 \\ & i,j=1,2,\ldots N \\ \end{align}\]

In Eq. (25), Doc1 = (TWP₁,TWP₂,…TWP_N) and Doc2 = (TWQ₁,TWQ₂,…TWQ_N), the weight values of the subject in the text are denoted by TWP_i and TWQ_j. The amount of transfer from TWP_i to TWQ_j is denoted by tranT(i,j). From Eq. (25), we get consT(Top_i,Top_j). From this, the text distance metric formula W2v_dist based on EMD distance and LDA topic model is obtained.

The Firefly algorithm is simple in structure, robust, easy to implement, and strong in finding ability. Therefore, this paper applies the Firefly algorithm to text clustering. Next, this paper will describe in detail the text clustering model based on the firefly algorithm.

Text clustering is first to do preprocessing operations on the text to remove the deactivated words. Then, using feature selection or feature extraction, the best word items are selected to express the text features. Finally, its most fundamental process is realized through cluster analysis.Thus, the Firefly algorithm will be used for the final stage of clustering implementation.

The current research application is the VSM, which converts textual semantic associations into spatial modeling similarities. These similarities can then be made available to clustering algorithms for manipulation.After obtaining the feature words from the text dataset through feature selection or extraction, each document can be represented in the following form: (26) di=(w1,w2,…,wn) ${{d}_{i}}=\left( {{w}_{1}},{{w}_{2}},\ldots ,{{w}_{n}} \right)$

In the text clustering process, each text is composed of feature vectors, where each dimension represents the weight of the corresponding feature item in this piece of data. Two n-dimensional objects i and j are known and their text feature vectors are denoted as: i = (x_i1,x_i2,…,x_in), j = (x_j1,x_j2,…,x_jn). The measures of similarity of text i and text j are:

1) Euclidean distance: (27) sim(i,j)=d(i,j)=| xi1−xj1 |2+| xi2−xj2 |2+⋯+| xin−xjn |2 \[sim\left( i,j \right)=d\left( i,j \right)=\sqrt{{{\left| {{x}_{i1}}-{{x}_{j1}} \right|}^{2}}+{{\left| {{x}_{i2}}-{{x}_{j2}} \right|}^{2}}+\cdots +{{\left| {{x}_{in}}-{{x}_{jn}} \right|}^{2}}}\]

From the above analysis, it can be learned that during the flight of the firefly, the step size of its movement has a direct impact on the performance of the algorithm. Therefore, it is very critical to set an appropriate position update strategy.

This paper utilizes the FA intelligent bionic algorithm to find the optimal solution ability, fast convergence speed and other characteristics, at the same time, for the traditional FA algorithm deficiencies made improvements, and the improved firefly algorithm AFA combined with the idea of the K-medoids algorithm is applied to the text clustering, and a new firefly clustering algorithm (K-AFA) is proposed. This paper describes three aspects: the selection of the objective function, the idea of the algorithm, and the process of the algorithm.

FA algorithm is based on the brightness of each firefly to search for the optimal solution of the search, usually set the brightness of the firefly as the value of the objective function, so the selection of the objective function directly affects the final algorithm results. In the K-center point algorithm, the point with the smallest arithmetic mean sum of the data to the rest of the data in the cluster is selected as the center of the clustering. So the objective function of this algorithm is defined as follows: (32) f(i)=1Nk∑j∈Skdij \[f\left( i \right)=\frac{1}{{{N}_{k}}}\sum\limits_{j\in {{S}_{k}}}{{{d}_{ij}}}\] Where N_k denotes the number of data elements in the cluster S_k where data element i is located, and d_ij denotes the distance between data element i and data element j, generally the Euclidean distance, which is judged by the cosine angle between documents in text clustering.

The firefly FA algorithm is applied to the problem of selection of clustering centroids, in general, the FA algorithm is to search for the individual with the largest brightness, while the clustering centroid selection is to search for the point that minimizes the value of the objective function when the number of clusters in the cluster is given a value, for this reason, the brightness of the firefly s is defined as: (33) I(s)=1f(s) \[I\left( s \right)=\frac{1}{f\left( s \right)}\]

The higher the value of I(s), the smaller the value of the function denoting f(s), then the lower the value of the objective function sought, the closer the optimal solution obtained in the end is to the ideal solution, and the better the final clustering effect.

Using the Firefly text theme clustering algorithm constructed in this paper to analyze the specific publication year of 210 papers, it can be found that: since the 1930s, women’s consciousness research began to revive, and the research results showed a general increasing trend, the number of papers increased rapidly during 1945, and after 1949 there was even a surge in the number of papers. The yearly values are shown in Figure 2.

Figure 2.

The paper publishes the annual score

Select 179 of them and study their research themes and contents.

The statistics of themes and contents of women’s consciousness research are shown in Figure 3. The themes of women’s consciousness research since the Anti-Japanese War have been widely distributed, but unevenly. Among the eight categories of themes and contents summarized, background research on the meaning, value, and practical foundation of cultivating women’s consciousness takes the first place, accounting for 49.721% of the total research, research on the status quo, problems, and countermeasures of cultivating women’s consciousness takes the second place, accounting for about 14.525% of the total research, and research on the social hotspots and phenomena caused by the lack of women’s consciousness accounts for 13.966% of the total research, and research involving Women, the concept, connotation and characteristics of women’s consciousness have not been given due attention, accounting for only 10.056% of the total. Historical studies on the development of women’s consciousness, comparative studies on theories, ideologies, policies and experiences of women’s consciousness in foreign countries as well as trends in the development of women’s education in the context of globalization, and the measurement and evaluation of women’s consciousness as well as a review of women’s consciousness have received less attention from the scholars. Measurement and evaluation of women’s consciousness and review of women’s consciousness studies have received less attention from scholars.

Figure 3.

Women’s consciousness research theme and content statistics

Figure 4 shows the distribution of topics, content, and methods related to women’s consciousness research. The field of women’s consciousness research in China mainly adopts qualitative research methods, and among the 179 samples in this statistical survey, a total of 130 articles, or about 72.626%, have been used in literature analysis and theoretical discursive research. In contrast, comparative studies, case studies, experimental studies, and multivariate studies are rarely used. As an indispensable research method in scientific research, theoretical discursive research is of great significance to the construction of the basic research system of women’s consciousness, however, as an important practical field, the cultivation of women’s consciousness, its effectiveness, the current situation of women’s consciousness and the important factors influencing it at the micro level need to be supported by scientific research and empirical analysis, so as to make women’s consciousness practice work in a targeted way. However, as an important field of practice, the cultivation of women’s consciousness needs scientific research and empirical analysis to support its effectiveness, the current situation of women’s consciousness and the important factors influencing it at the micro level.

Figure 4.

Women’s consciousness research topics and content, method distribution

Special groups such as farmers and migrant workers, primary and secondary school students, and university students are gradually receiving attention. Figure 5 shows the distribution of research groups on women’s consciousness, with the largest number of researches on the ideal general group, accounting for almost 58.659% of the total researches, and the special groups of citizens, farmers and migrant workers, college students, primary and secondary school students, party and governmental organs and the military, ethnic minority groups, and enterprise units have all been involved, among which the college students’ group has received a higher degree of attention. With the spread of the construction of the new socialist countryside and the prominence of the problem of rural migrant workers, the study of the female consciousness of farmers and rural migrant workers has begun to attract the attention of the academic world. In contrast, women’s awareness in primary and secondary schools, which is the main foundation of women’s education in China, has not received enough attention.The study of women’s consciousness in public, party, government, and military organizations, ethnic minority groups, and enterprise groups has been favored by only a few scholars and needs further attention from scholars.

Figure 5.

Women’s consciousness research group distribution

The variables were categorized into explanatory, interpretive, and control variables according to the purpose of the study. The assigned values and descriptive statistics of each variable are shown in Table 1.

In the main effect analysis of the influence of era background and social environment, cultural values on the awakening of women’s consciousness, as shown in Table 2, Model 1, Model 2 and Model 3 are the regression results obtained from the fitting of ordered Logit model, and Model 4 is the regression results obtained from the fitting of multiple linear regression model. Among them, Model 1 is the result of regression analysis with only control variables, age, political appearance, marital status, years of education, type of household registration, type of region, health status, and family economic status all significantly and positively affect women’s awakening of consciousness, while work status does not show significant statistical significance.

The household registration system is an important feature of the urban-rural dichotomy, and the type of household registration causes differences in resource endowments, lifestyles, and social attitudes among different groups, which significantly affects the group’s class identity. This study further conducted a sub-sample regression analysis of women’s awakening of consciousness based on household registration type, and Table 3 shows the analysis of urban-rural differences in the influence of era background and social environment, and cultural values on women’s awakening of consciousness. It is found that there is a significant difference between urban and rural areas in the influence of contemporary background and social environment on women’s awakening to consciousness, while there is no significant difference between urban and rural areas in the influence of cultural values on women’s awakening to consciousness.

Both horizontal and vertical socio-economic status comparisons show that socio-economic status significantly and positively affects the class identity of rural and urban women at the 1% level, further validating hypothesis H2. Meanwhile, the adjusted R² is 0.115 and 0.189 in rural and urban areas, respectively, with urban areas being more awakened to women’s consciousness than rural areas.