Network reputation evaluation method and system fusing emotional features and semantic features

Abstract

The invention discloses a network reputation evaluation method and system fusing emotional features and semantic features. The method comprises the steps that text data related to an evaluation object is collected and preprocessed; completing extraction of emotional special diagnosis and semantic special diagnosis; single-dimensional evaluation is carried out on three dimensions of emotional influence, theme influence and behavior influence; dimensionality reduction processing is carried out on evaluation indexes of all dimensions such as emotion influence, theme influence and behavior influence, principal components and factors are extracted, and an evaluation model is simplified; dividing the network entities into different categories through hierarchical clustering analysis, and performing network reputation evaluation according to features of different categories; evaluating the efficiency of network reputation management from the perspective of input and output based on a DEA efficiency model; and calculating a weight matrix of the evaluation indexes based on the BP neural network, and carrying out prediction evaluation on the network reputation. The invention provides the network reputation evaluation method and system fusing the emotion features and the semantic features, and the comprehensive, scientific and accurate evaluation of the network reputation is realized.

Description

Technical Field

[0001] The present invention relates to the field of network reputation evaluation, and specifically to a method and system for network reputation evaluation by combining emotional features and semantic features. Background Art

[0002] In today's digital information age, the rapid development of the Internet has made the cyberspace an important place for people to communicate, interact, spread information and generate public opinion. As the image and word-of-mouth of a subject in the network environment, network reputation has a crucial impact on its market competitiveness, social influence and sustainable development ability. A good network reputation can attract more resources and enhance public trust, while a negative network reputation may lead to customer loss, damaged brand image, and even cause serious economic and social consequences. Therefore, accurately and effectively evaluating network reputation has become an important issue of wide concern in all fields.

[0003] However, the existing network reputation evaluation methods have many limitations. First, the evaluation dimension is single. Most existing methods only evaluate network reputation from a single specific angle. For example, they simply rely on sentiment analysis to judge the positive or negative sentiment tendency of text, or only measure network influence based on user behavior data (such as the number of likes, comments, forwards, etc.). This single-dimensional evaluation method cannot comprehensively and deeply reflect the complex connotation and multi-faceted characteristics of network reputation, and is prone to one-sidedness and inaccuracy of the evaluation results. Second, the feature fusion is insufficient. In network text data, emotional features and semantic features are two key aspects that are interrelated and have important impacts on network reputation evaluation. However, current evaluation methods often analyze them separately, failing to fully explore the internal connection and synergistic effect between them, thus making it difficult to accurately grasp the true attitudes and intentions of users and unable to conduct a comprehensive and integrated evaluation of network reputation. Third, the data processing and analysis capabilities are limited. Network text data is characterized by being massive, diverse, complex and dynamically changing. Existing technologies have many difficulties in data preprocessing, feature extraction, model construction and training, such as being difficult to effectively remove noise data, accurately extract key features, and construct a stable and reliable evaluation model, resulting in the reliability and effectiveness of the evaluation results being affected.

[0004] In summary, the existing network reputation evaluation methods cannot meet the needs of practical applications, and there is an urgent need for an innovative network reputation evaluation method and system that can integrate emotional features and semantic features to overcome the defects of the existing technologies and achieve a comprehensive, scientific and accurate evaluation of network reputation. The present invention comes into being under such a background. Summary of the Invention

[0005] The present invention constructs a network reputation evaluation method and system that integrates emotional features and semantic features, solving the problems existing in the existing network reputation evaluation methods from multiple aspects; at the same time, the prediction and evaluation function based on the BP neural network can also provide a forward-looking reference for decision-making, helping to take measures in advance to cope with potential reputation crises.

[0006] In view of the above deficiencies or improvement requirements of the prior art, the present invention provides a network reputation evaluation method that integrates emotional features and semantic features, including:

[0007] S1. Collect text data related to the evaluation object, clean the original text, segment the text into words or phrases, confirm stop words, and perform noise reduction processing;

[0008] S2. Adopt an emotional analysis method, combined with the cosine similarity algorithm, to perform an emotional tendency analysis on the preprocessed text data;

[0009] S3. For the preprocessed text data, convert the high-frequency keywords in the text into word vectors through the word2vec model to realize the vector representation of text semantics;

[0010] S4. Perform single-dimensional evaluations in three dimensions: emotional influence, theme influence, and behavior influence;

[0011] S5. Integrate emotional features and semantic features for multi-dimensional index fusion comprehensive evaluation. The specific steps are as follows:

[0012] Perform dimensionality reduction processing on the evaluation indexes of each dimension such as emotional influence, theme influence, and behavior influence, extract the principal components and factors, and simplify the evaluation model;

[0013] Through hierarchical clustering analysis, divide network entities into different categories, and perform network reputation evaluation according to the characteristics of different categories;

[0014] Evaluate the efficiency of network reputation management from the perspectives of input and output based on the DEA efficiency model;

[0015] Based on the BP neural network, calculate the weight matrix of the evaluation indexes and perform a prediction and evaluation of the network reputation.

[0016] Furthermore, the emotional analysis method in S2 is as follows:

[0017] Judge whether the words in the text are emotional words according to the emotional dictionary, obtain the emotional degree words and negative words, determine the emotional polarity of the words, and then judge the emotional polarity of the sentence, and calculate the emotional tendency of the document; further calculate the emotional value similarity of the mutually related documents according to the obtained emotional tendency value;

[0018] The specific method for calculating the emotional tendency of the document is as follows:

[0019] S θ = α / β * 100%

[0020] Where α is the number of sentences with a certain type of sentiment polarity, β is the number of sentences in document S, and the proportion of a certain type of polarity sentiment is the value of a certain type of polarity tendency of a document S θ 。

[0021] Furthermore, the word2vec model in S3 realizes the vectorization processing of text vocabulary through Skip-gram, specifically as follows:

[0022] The basic Skip-gram model is defined as p(w i | w t ); Let w be the input vector, V w be the output vector, the size of the dictionary be N, and the probability of correctly predicting w j through the keyword w i is as follows:

[0023]

[0024] In the formula, p(w i | w t ) represents the probability of the occurrence of word w t under the condition of the given word w i ; represents the transpose of vector V i . In mathematics and linear algebra, the transpose of a vector is an operation that converts a row vector into a column vector or a column vector into a row vector; in the context of the Skip-gram model, usually represents the vector representation of the input vector w t , while represents the vector representation of the output vector W i . The expression represents the dot product (inner product) of these two vectors, that is, the sum of the products of the corresponding elements, and is used to calculate the similarity between the input vector and the output vector.

[0025] Furthermore, in S4, the single-dimensional evaluation is carried out separately in three dimensions of sentiment influence, theme influence, and behavior influence. The specific methods are as follows:

[0026] Sentiment influence evaluation: Based on sentiment analysis and cosine similarity algorithm, calculate the sentiment tendency value and sentiment value similarity of the text, and evaluate the sentiment influence of the network reputation;

[0027] Theme influence evaluation: Use word2vec word vector processing technology and cosine similarity algorithm to calculate the theme similarity of the text, and evaluate the theme influence of the network reputation;

[0028] Behavior impact evaluation: By statistically analyzing user behavior data such as likes, comments, and forwards, evaluate the behavior impact of online reputation.

[0029] Further, the specific method for simplifying the evaluation model in S5 is as follows:

[0030] Perform principal component analysis on the preprocessed small sample data set and calculate the covariance matrix:

[0031]

[0032] Perform eigenvalue decomposition on C:

[0033] Cv k =λ k v k

[0034] where λ k is the eigenvalue and v k is the corresponding eigenvector

[0035] Sort the eigenvalues from largest to smallest, and select the eigenvectors corresponding to the top p largest eigenvalues to form the transformation matrix V = [v1, v2, …, v p ,];

[0036] Through:

[0037] Y = ZV

[0038] Project the original data into the new low-dimensional space to obtain the principal component matrix Y, whose dimension is m × p. These principal components are the key factors affecting online reputation;

[0039] Suppose Z can be expressed as Z = AF + ∈, where A is the factor loading matrix, F is the common factor matrix, and ∈ is the special factor matrix. Estimate the factor loading matrix A and the common factor matrix F through methods such as maximum likelihood estimation; select the common factors with larger variance contribution rates and combine them with the principal component analysis results to obtain the key evaluation factor set K, providing a basis for subsequent cluster analysis.

[0040] Further, the specific method for dividing network entities into different categories through hierarchical cluster analysis and evaluating online reputation according to the characteristics of different categories is as follows:

[0041] Based on the key evaluation factor set K, calculate the distance between samples using the Euclidean distance. For two samples k i and k j , their Euclidean distance:

[0042]

[0043] where and are the sample and the value on the l-th key evaluation factor respectively;

[0044] The agglomerative hierarchical clustering algorithm is adopted. Initially, each sample point is regarded as a separate class. In each step of the merging process, the distances between different classes are calculated. The minimum distance method is used, that is, the distance between two classes is the distance between the two closest sample points in the two classes;

[0045] Let class C a and class C b ,

[0046]

[0047] Merge the two closest classes until all sample points are merged to reach the preset number of clusters, and obtain the set of reputation classes C clustered after clustering, providing an intuitive reference for efficiency evaluation.

[0048] Furthermore, the specific method for evaluating the efficiency of network reputation management from the perspectives of input and output based on the DEA efficiency model is as follows:

[0049] Suppose there are q decision-making units, that is, network entities. The input index is the matrix X composed of all evaluation influencing factors, with a dimension of q×p, where p is the number of key evaluation factors, and the output index is the matrix Y composed of key indexes representing network reputation output , with a dimension of q×r, where r is the number of output indexes.

[0050] Adopt the CCR model based on constant returns to scale. For the i-th decision-making unit, construct a linear programming problem:

[0051]

[0052] λ j ≥0, j = 1, …, q

[0053] The objective function is maxθ i , where θ i represents the efficiency value of the i-th decision-making unit; the DEA model aims to adjust the weight variable λ j to maximize the efficiency value θ i of decision-making unit i; in the scenario of network reputation evaluation, this means finding an optimal weight allocation method to evaluate the efficiency of this network entity in using input resources to obtain output;

[0054] Input constraint: The constraint condition means that for the r-th input index, the total input weighted by all decision-making units according to the weight λ j shall not exceed the actual input value x of the i-th decision-making unitir ; where q is the total number of decision-making units, and λ j is the weight variable of the j-th decision-making unit, and x jr is the value of the r-th input index of the j-th decision-making unit;

[0055] Output constraint: The constraint condition indicates that for the s-th output index, the total output of all decision-making units weighted by λ j should be greater than or equal to the actual output value y of the i-th decision-making unit is multiplied by its efficiency value θ i ; where y js is the value of the s-th output index of the j-th decision-making unit;

[0056] Non-negative weight constraint: The constraint condition λ j ≥0, j = 1, …, q ensures that the weight variable λ j is non-negative; in practical terms, the weight represents the contribution degree of other decision-making units to the efficiency evaluation of the i-th decision-making unit, and non-negative weights are reasonable because negative contributions cannot occur;

[0057] By solving the above linear programming problem, the efficiency value θ of each decision-making unit can be obtained i . The closer the efficiency value is to 1, the higher the efficiency of the decision-making unit in using input resources to obtain outputs such as network reputation, thus providing a strong basis for screening key influencing indicators and optimizing network reputation management strategies.

[0058] Furthermore, the specific method for predicting and evaluating network reputation by calculating the weight matrix of evaluation indicators based on the BP neural network is as follows:

[0059] Model input: All evaluation indicators based on a large-scale dataset form the input matrix I large , with a dimension of M×n, where M is the number of samples in the large-scale dataset;

[0060] Model training: Using the principal component and factor analysis results under small-sample data, that is, the set of key evaluation factors as the target output;

[0061] Suppose the BP neural network includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is n, the number of nodes in the hidden layer is h, and the number of nodes in the output layer is p;

[0062] The weight matrix from the input layer to the hidden layer is W ih , and the weight matrix from the hidden layer to the output layer is W ho ;

[0063] The input of the hidden layer nodes after being processed by the activation function σ, obtains the hidden layer output h out= σ(h ih );

[0064] The input of the output layer node The predicted output is

[0065] Through the backpropagation algorithm, calculate the predicted value The error between the predicted value and the target value k:

[0066]

[0067] Adjust the weights according to the error and continuously iterate for training.

[0068] As the second aspect of the present invention, there is provided a network reputation evaluation system that fuses emotional features and semantic features, including:

[0069] A data collection and preprocessing unit for collecting text data related to the evaluation object, cleaning the original text, splitting the text into words or phrases, identifying stop words, and performing noise reduction processing;

[0070] An emotional feature extraction unit for performing emotional tendency analysis on the preprocessed text data by using an emotional analysis method in combination with a cosine similarity algorithm;

[0071] Semantic feature extraction, which is used to convert high-frequency keywords in the text into word vectors through a word2vec model for the preprocessed text data, and realize the vector representation of text semantics;

[0072] A single-dimensional evaluation unit for performing single-dimensional evaluations in three dimensions of emotional influence, theme influence, and behavior influence respectively;

[0073] A multi-dimensional fusion comprehensive evaluation unit for performing multi-dimensional comprehensive evaluations by fusing emotional features and semantic features. The specific steps are as follows:

[0074] Perform dimensionality reduction processing on the evaluation indicators in each dimension such as emotional influence, theme influence, and behavior influence, extract the main components and factors, and simplify the evaluation model;

[0075] Through hierarchical clustering analysis, divide network entities into different categories, and perform network reputation evaluations according to the characteristics of different categories;

[0076] Evaluate the efficiency of network reputation management from the perspectives of input and output based on the DEA efficiency model;

[0077] Based on the BP neural network, calculate the weight matrix of the evaluation indicators and perform a predictive evaluation of the network reputation.

[0078] As a third aspect of the present invention, there is also provided a computer-readable storage medium, on which a computer program is stored, characterized in that the computer program is executed by a processor to perform any step of the above-mentioned network reputation evaluation method that integrates emotional features and semantic features.

[0079] Generally speaking, compared with the prior art by the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

[0080] 1. The network reputation evaluation method that integrates emotional features and semantic features of the present invention evaluates the network reputation by integrating emotional and semantic features, and comprehensively uses principal component, factor, cluster analysis, DEA model and BP neural network. The perspective of traditional single evaluation methods is limited, which is likely to lead to deviation of results. The present invention combines multiple methods. Principal component and factor analysis are used for dimensionality reduction to remove redundant information; cluster analysis is used for classification evaluation; DEA model is used to measure efficiency; BP neural network is used for accurate modeling. Multiple measures are taken to analyze the reputation from different angles, overcome the limitations of single methods, comprehensively consider various factors, significantly improve the accuracy and reliability of evaluation results, and provide a solid basis for decision-making.

[0081] 2. The network reputation evaluation method that integrates emotional features and semantic features of the present invention integrates emotional, semantic and user behavior features. Single feature evaluation is one-sided and cannot comprehensively reflect the reputation. Emotional features reflect subjective attitudes, semantic features reveal core content, and behavior features reflect actual actions. The integration of the three complements each other's advantages and comprehensively considers from multiple dimensions. It can capture network reputation information in all aspects, overcome the limitations of single evaluation, avoid information loss, and more comprehensively and truly reflect the network reputation situation, helping managers accurately grasp the reputation situation.

[0082] 3. The network reputation evaluation method that integrates emotional features and semantic features of the present invention realizes dynamic prediction of network reputation based on BP neural network. Traditional evaluations are mostly static and difficult to cope with the rapidly changing network environment. The BP neural network can learn historical data and multi-dimensional indicators to capture the changing rules of reputation. Based on this, the future reputation trend is predicted, providing a forward-looking reference for decision-making. Managers can detect potential risks in advance, formulate strategies in advance, take measures to avoid damage to the reputation, and enhance the initiative and effectiveness in dealing with reputation crises. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 is a flowchart of the network reputation evaluation method that integrates emotional features and semantic features according to an embodiment of the present invention;

[0084] Figure 2 is a schematic diagram of the analysis path of the emotional impact evaluation index according to an embodiment of the present invention;

[0085] Figure 3 is a schematic diagram of the difference between Euclidean distance and cosine similarity according to an embodiment of the present invention;

[0086] Figure 4 Schematic diagram of the analysis path of the theme impact evaluation index for the embodiment of the present invention;

[0087] Figure 5 Schematic diagram of the Skip-Gram model architecture for the embodiment of the present invention;

[0088] Figure 6 Schematic diagram of the comprehensive system for the embodiment of the present invention;

[0089] Figure 7 Schematic diagram of the system unit for the embodiment of the present invention. Detailed implementation manners

[0090] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0091] Embodiment 1

[0092] Please refer to Figure 1 , Embodiment 1 of the present invention provides a network reputation evaluation method that integrates emotional features and semantic features, including:

[0093] S1. Collect text data related to the evaluation object, clean the original text, segment the text into words or phrases, identify stop words, and perform noise reduction processing;

[0094] S2. Determine whether the words in the text are emotional words according to the emotion dictionary, obtain the degree of emotion words and negative words, determine the emotional polarity of the words, and then judge the emotional polarity of the sentence, calculate the emotional tendency of the document; further calculate the similarity of the emotional values of the mutually related documents according to the obtained emotional tendency values;

[0095] S3. For the preprocessed text data, convert the high-frequency keywords in the text into word vectors through the word2vec model to realize the vector representation of the text semantics;

[0096] S4. Perform single-dimensional evaluation in three dimensions of emotional impact, theme impact, and behavioral impact respectively;

[0097] S5. Integrate emotional features and semantic features for multi-dimensional index fusion comprehensive evaluation. The specific steps are as follows:

[0098] Perform dimensionality reduction processing on the evaluation indexes of each dimension such as emotional impact, theme impact, and behavioral impact, extract the main components and factors, and simplify the evaluation model;

[0099] Through hierarchical clustering analysis, network entities are divided into different categories, and network reputation evaluation is carried out according to the characteristics of different categories; from the perspectives of input and output, the efficiency of network reputation management is evaluated based on the DEA efficiency model; the weight matrix of evaluation indicators is calculated based on the BP neural network to predict and evaluate network reputation.

[0100] The implementation steps of the present invention are specifically elaborated as follows:

[0101] The purpose of this Embodiment 1 is to provide a network reputation evaluation method that integrates emotional features and semantic features. This method includes data preprocessing, emotional feature extraction, semantic feature extraction, and comprehensive evaluation.

[0102] (1) Data preprocessing

[0103] Data preprocessing is responsible for performing a series of processes on the collected original text data, including steps such as data cleaning, word segmentation, and stop word removal.

[0104] Data cleaning: Remove irrelevant characters, special symbols, etc. in the original text to ensure the purity of the text data.

[0105] Word segmentation: Use a word segmentation tool to split the text data into words or phrases, laying a foundation for subsequent emotional feature extraction and semantic feature extraction.

[0106] Stop word removal: Remove common but meaningless stop words in the text, such as "de" (of), "le" (particle indicating completion), etc., to reduce noise interference.

[0107] In addition, the preprocessing module can also perform part-of-speech tagging, syntactic analysis, etc. on the text to facilitate more accurate feature extraction by subsequent modules.

[0108] The emotional feature extraction module uses an emotional analysis method and combines the cosine similarity algorithm to perform an emotional tendency analysis on the preprocessed text data. This module can identify emotional words in the text and judge their emotional tendencies (positive, negative, or neutral) according to the context. At the same time, the cosine similarity algorithm is used to calculate the similarity between the text and the words in the emotional dictionary to further determine the emotional tendency and its intensity of the text.

[0109] The semantic feature extraction module uses a semantic analysis method and combines the word2vec word vector processing technology to perform semantic analysis on the preprocessed text data. This module can identify keywords and themes in the text and represent the text as a vector form using word2vec word vectors for subsequent similarity calculation and clustering analysis. In addition, this module can also use methods such as topic models to extract potential theme information in the text to enrich semantic features.

[0110] The comprehensive evaluation module fuses the features obtained by the sentiment feature extraction module and the semantic feature extraction module, and comprehensively evaluates the network reputation using a variety of evaluation methods. This module can use a method combining principal component and factor analysis to reduce the dimensions of the network reputation index for each dimension, so as to reduce information overlap and endogeneity problems. At the same time, the hierarchical clustering method is used to divide network entities into different categories, and the network reputation is evaluated according to the characteristics of different categories. In addition, this module can also perform prediction evaluation and efficiency evaluation based on methods such as the DEA efficiency model and BP neural network, providing rich reference information for the scientific and reasonable evaluation of network reputation.

[0111] (2) Sentiment feature extraction

[0112] Please refer to Figure 2 , and on the basis of the core evaluation target of sentiment influence, the parsing path of sentiment influence characterization indicators as shown in Figure 2 is proposed. After completing the preliminary basic work such as determining samples, data sources, and collection methods, it is necessary to preprocess the text, such as word segmentation and noise reduction, then calculate the sentiment tendency value of tweets and the sentiment tendency value of comments respectively, and finally calculate the similarity to detect the sentiment influence of network topics.

[0113] Among them, data preprocessing mainly includes the following two steps:

[0114] Chinese word segmentation. The purpose of word segmentation is to split a sentence or a whole paragraph of text into independent words. Based on the reality of text mining, the most widely used method at present is the statistical analysis method. This method can master the word segmentation rules through continuous training and self-learning, and apply them to new texts. In this paper, text data published and comment data of the discourse audience are collected through Python, and the Jieba word segmentation module that can be directly called by Python is also selected as the word segmentation tool. The biggest feature of this module is its high flexibility. Users can add custom dictionaries to formulate word segmentation rules for different corpora. This paper deals with Sina Weibo data. In view of the characteristics of text spoken language and networking, a custom dictionary for Weibo language is added to support word segmentation processing.

[0115] Noise reduction processing. After the text is segmented, there are still meaningless function words and content words, such as modal particles, online expressions, university names, etc. These words have no practical significance for analyzing sentiment tendency and discourse topics. In this paper, these words are collectively referred to as noise data. The existence of a large amount of noise data will affect the accuracy of sentiment analysis and topic training, and it is necessary to perform noise reduction processing on the document data.

[0116] First, stop words are processed. When performing sentiment analysis on text, it is necessary to extract feature words with sentiment tendencies. Usually, there are words with high frequencies in the text that are meaningless for extracting feature words, such as conjunctions, prepositions, symbols, and punctuation. Words of the above types cannot become the feature words of the document and are collectively referred to as stop words. In this paper, a stop word list is established. Through a self-written Python program, the words in the document are looped through and compared with the words in the stop word list, and the words that match successfully in the stop word list are deleted. The stop word list used in this paper includes a general stop word list, an extended version of the Harbin Institute of Technology stop word list, and a Google search engine stop word list. In addition, according to the characteristics of the online language on Sina Weibo, the stop word list is continuously added to and adjusted through the analysis of the result data.

[0117] Secondly, specific online languages are processed. The sentiment analysis text used in this paper is Sina Weibo text. As a short text-based social network platform, Sina Weibo has formed unique language characteristics different from general texts during its continuous development, such as emojis and symbols. Although these emojis and symbols have specific meanings in the context and semantic environment at that time, they are independent symbols in the bag-of-words structure, and it is difficult to judge their semantics through modeling. Therefore, this paper classifies them as noise data that needs to be removed. In addition, the symbols "@" and " / / " that often appear during the Weibo forwarding process also have no special meaning, and these special symbols need to be added to the stop word list for processing as noise data.

[0118] This paper selects a sentiment dictionary processing method based on sentence analysis to calculate the sentiment tendency of the text. The sentiment analysis dictionary selects the HowNet dictionary, which contains a conceptual knowledge base of Chinese and English words, and can reveal features such as lexical concepts, relationships between concepts, and attributes of concepts. The sentiment dictionary based on this knowledge base is mainly divided into two parts: Chinese and English, and contains rich Chinese and English evaluation words, sentiment words, etc.

[0119] This paper uses the Python self-programming language, and takes the tweet information and comment information of each Weibo account as a document respectively. Each document consists of several sentences, and each sentence consists of several words. First, it is judged whether the word is a sentiment word (that is, whether it is in the sentiment dictionary), and the sentiment intensity word and the negative word are obtained, and then the sentiment polarity (positive, negative, and neutral) of the word is judged. Finally, the sentiment polarity of the sentence where the word is located is judged through the sentiment polarity of the word, so as to calculate the proportion of sentences with three different sentiment polarities in each document to obtain the sentiment tendency of the document. The specific formula is as follows:

[0120] S θ = α / β * 100%

[0121] Among them, α is the number of sentences with a certain type of sentiment polarity, β is the number of sentences in document S, and the proportion of a certain type of polarity sentiment is the value of a certain type of polarity tendency S of a document θ 。

[0122] As shown in Table 1, through sentiment analysis, we can obtain the sentiment tendency values of each tweet set and each comment set. Next, we need to calculate the similarity between these two sets.

[0123] Table 1 Example table of sentiment tendencies of tweet sets and comment sets based on the Weibo platform

[0124]

[0125] There are many methods for calculating the difference or similarity between two individuals (X = χ1, χ2, χ3, … X n )(Y = y1, y2, y3, … y n ) with n-dimensional features. Among them, common ones include correlation analysis and clustering analysis. In addition, the similarity of sentiment vectors can also be analyzed through vector similarity measurement. Similarity measurement is characterized by the "distance" between samples. Common calculation methods include Euclidean Distance, Minkowski Distance, Manhattan Distance, Chebyshev Distance, Mahalanobis Distance, Similarity, cosine similarity, Pearson correlation coefficient, Jaccard Coefficient, etc. Euclidean distance and cosine similarity are two of the most basic and widely applicable calculation methods, and other methods are basically variations and derivatives of these two methods.

[0126] Euclidean distance, abbreviated as Euclidean distance, is mainly used to measure the absolute distance between two points in a multi-dimensional space. The formula is as follows:

[0127]

[0128] Cosine similarity measures the magnitude of the difference between individuals using the cosine value of the angle between two vectors in a multi-dimensional space. The formula is as follows:

[0129]

[0130] Euclidean distance is a distance similarity measurement, and cosine similarity is a direction similarity measurement. There are differences between the two, and the differences between the two can be seen more intuitively with the help of a three-dimensional coordinate system. Please refer toFigure 3 , Euclidean distance is mainly used to measure the absolute distance between two points, which is determined by the absolute values of each point and can reflect the differences in individual values, and is mostly used to compare the magnitudes of values. Cosine similarity mainly determines similarity by considering the angle between two vectors and is insensitive to absolute values, and is used to distinguish the differences between individuals. This paper attaches more importance to the consistency of emotional direction in the measurement of emotional impact, so the cosine similarity algorithm is selected to measure the emotional impact of network issues.

[0131] Cosine similarity cannot measure the differences in each dimension value, which will lead to a special evaluation error. The Adjusted Cosine Similarity algorithm corrects this unreasonable error by subtracting a mean value from all dimension values. For example, if the ratings of two users are (1, 2) and (4, 5) respectively, and their average rating is 3, then after adjustment, they are (-2, -1) and (1, 2), and then calculated using cosine similarity, the result is -0.8. The similarity is negative and the difference is large, which is more in line with reality. The calculation of the emotional similarity index in this paper is also based on the adjusted cosine similarity algorithm.

[0132] (3) Semantic feature extraction

[0133] Please refer to Figure 4 , this paper designs an analysis path for the evaluation index of topic influence in the dimension of discourse influence. Its focus is to measure the similarity between the topic of the information released on Weibo and the topic of the comments of the discourse audience, and the following measurement model is designed based on this idea.

[0134] The data sources of the topic similarity and emotional similarity indexes in this section are the same, but the measurement objects are different, and the data processing processes also have differences. First, it is also necessary to preprocess the Weibo dataset and the comment dataset, including word segmentation, noise reduction, etc. Secondly, the word2vec method is used to process the word vectors of 200 high-frequency keywords in each set. Then, the cosine similarity is used to measure the similarity of the word vectors of the two sets, so as to obtain the topic similarity index of each Weibo set and comment set. The calculation methods and principles of word segmentation, noise reduction and cosine similarity have been elaborated in detail above and will not be repeated here. This subsection mainly introduces the word2vec word vector processing method.

[0135] The word2vec model was first proposed by Mikolov et al. Its main function is to convert words into word vectors and has been gradually widely applied in natural language processing. The word2vec model is a simplified neural network. The input layer is a One-Hot Vector, and there is no activation function in the hidden layer, that is, a linear unit. The Softmax regression function method is used between the output layer and the input layer. After training with the word2vec model, it can provide vector representations for calculating the semantic similarity of texts. Word2vec generally realizes the vectorization processing of text vocabulary through two architectures: CBOW (Continuous Bag-of-Words) and Skip-gram. Among them, the CBOW model outputs the word vector of the feature word through the word vectors of the context, while Skip-Gram outputs the context word vectors according to the word vector of a specific word. In this study, the word2Vec toolkit provided by the gensim module in Python was used for training to realize the word vector processing of network discourse content and comment content, and its principle is the Skip-gram model.

[0136] Please refer to Figure 5 , the Skip-gram model mainly predicts the probability p(w i |w t ), where t - c ≤ i ≤ t + c, c represents the parameter determining the context window. The larger c is, the more context is considered, which will increase the training time while improving the accuracy. Therefore, reasonable parameters need to be set. w t is the current word. For a given word sequence w1, w2,..., w T , the goal of Skip-gram is to maximize the objective function, and the objective function is as follows:

[0137]

[0138] The basic Skip-gram model is defined as p(w i |w t ). Let w be the input vector, V w be the output vector, and the size of the dictionary be N. The probability of correctly predicting w j for the keyword w i is shown in the following formula.

[0139]

[0140] In the formula, p(w i |w t ) represents the probability of the occurrence of word w t under the condition of the given word w i ; represents the vector V iIn mathematics and linear algebra, the transpose of a vector is an operation that converts a row vector into a column vector, or a column vector into a row vector; in the context of the Skip-gram model, usually represents the vector representation of the input vector w t while represents the vector representation of the output vector W i . The expression represents the dot product (inner product) of these two vectors, that is, the sum of the products of the corresponding elements, which is used to calculate the similarity between the input vector and the output vector.

[0141] The objective function is trained by stochastic gradient descent to obtain the context vector based on Hierarchical Softmax representation. Based on the above word vector processing, this paper uses the cosine similarity algorithm to calculate the semantic similarity between text sets, and obtains the theme influence index of network topics. The cosine similarity algorithm has been elaborated in detail above and will not be repeated here.

[0142] (4) Comprehensive evaluation

[0143] Please refer to Figure 6 . By sorting out the above specific evaluation indicators, it is found that the current evaluation research mainly focuses on the management process or the communication process, and selects a certain angle to evaluate the communication power, influence or public opinion guidance. At present, there are few evaluation studies on the public opinion guidance of public opinion events. Most of them are studied from the front-end perspectives such as public opinion security assessment and public opinion monitoring assessment; while the research on the response management effect in public opinion events mainly starts from the perspective of government public opinion governance. Based on the high network reputation cognition and related theories, this paper selects relevant indicators with high relevance to evaluate the network reputation by combining the evaluation dimensions and the connotations of the constituent elements. Combining expert opinions, this paper initially selects 20 evaluation indicators that are highly relevant and relatively mature to the constituent elements of the network reputation evaluation dimension. The third-level indicators included in each dimension, the explanations of each indicator and the main reference sources are shown in Table 2.

[0144] Table 2 Evaluation indicators related to the constituent elements of the network reputation evaluation dimension

[0145]

[0146] Then, the quantification and methods of evaluation indicators were designed. The emotional impact indicator was calculated using sentiment analysis methods and cosine similarity methods. The theme impact indicator was obtained by combining the word2vec word vector processing method and the cosine similarity algorithm. The evaluation indicator of the negative public opinion crisis response ability was obtained by experts' scoring based on the response and feedback effects in negative hot events. The behavior impact indicator was obtained through the big data collection and analysis of the audience's behaviors such as liking and forwarding. The positive public opinion hot spot productivity indicator was extracted from the Weibo hot search data by analyzing positive public opinions.

[0147] Finally, an evaluation method integrating multiple methods was designed. Descriptive statistical analysis and correlation analysis were used for the data feature analysis of network reputation evaluation indicators. On the basis of deeply understanding the strength of the relationships between the internal and external measurement indicators and their mutual influences, it promoted the understanding of the formation process and operation mechanism of network reputation. The principal component and factor analysis were combined to reduce the dimensions of the indicators in each dimension of network reputation to solve the endogenous problems such as information overlap, and single-dimensional evaluation was carried out to more clearly understand the situation of each dimension. When comprehensively evaluating the network reputation, the hierarchical clustering method was used to divide it into different categories, and the results were mapped to the three-dimensional network space to realize the analysis of the characteristics of different categories. Based on the DEA efficiency model, the efficiency between the influencing factors and the characterizing factors of new media discourse power was evaluated from the perspective of input and output. Based on the BP neural network, the weight matrix of evaluation indicators was calculated to realize predictive evaluation, providing a reference idea for big data processing and prediction. The above evaluation methods depicted and described the network reputation in detail from multiple aspects, promoted the scientific and reasonable evaluation of network reputation in the new media environment, and provided richer reference information for decision-making.

[0148] In the evaluation of network reputation, a single evaluation method often fails to comprehensively reflect the true situation of network reputation. To overcome the limitations of existing evaluation methods, the present invention proposes an evaluation method integrating multiple methods. This method comprehensively uses a variety of statistical analysis and machine learning techniques to conduct multi-dimensional and multi-level comprehensive evaluation of network reputation, aiming to provide more scientific and accurate evaluation results and richer reference information for decision-making.

[0149] The framework of the evaluation method integrating multiple methods of the present invention includes the following key steps:

[0150] (1) Data preprocessing and feature extraction

[0151] Data preprocessing: Operations such as cleaning, word segmentation, and stop word removal were performed on the collected network text data to ensure the purity and usability of the data.

[0152] Feature extraction: Combining emotional features and semantic features, key features such as the emotional tendency value, theme similarity, and emotional value similarity of the text were extracted.

[0153] (2) Single - dimension evaluation

[0154] Emotional impact evaluation: Based on sentiment analysis and cosine similarity algorithm, calculate the sentiment tendency value and sentiment value similarity of the text, and evaluate the emotional impact of the online reputation.

[0155] Thematic impact evaluation: Use word2vec word vector processing technology and cosine similarity algorithm to calculate the thematic similarity of the text, and evaluate the thematic impact of the online reputation.

[0156] Behavioral impact evaluation: Evaluate the behavioral impact of the online reputation by statistically analyzing user behavior data such as likes, comments, and forwards.

[0157] (3) Multi - dimension comprehensive evaluation

[0158] Principal component analysis and factor analysis: Perform dimensionality reduction on the evaluation indicators of each dimension, extract the principal components and factors, reduce information overlap and endogeneity problems, and simplify the evaluation model.

[0159] Hierarchical clustering analysis: Divide network entities into different categories, evaluate the online reputation according to the characteristics of different categories, and realize the differential analysis of the online reputation of different categories.

[0160] DEA efficiency model: Evaluate the efficiency between the influencing factors and the characterizing factors of the online reputation from the perspective of input - output, and identify the efficiency bottlenecks in online reputation management.

[0161] BP neural network: Based on the BP neural network, calculate the weight matrix of the evaluation indicators, realize the predictive evaluation of the online reputation, and provide a forward - looking reference for decision - making.

[0162] Detailed steps for model construction

[0163] 1. Data pre - processing stage

[0164] Input data: The original online reputation dataset, covering multi - dimension evaluation indicators.

[0165] Processing process: Implement data cleaning, missing value filling, outlier detection and processing, and data standardization operations to ensure the data quality and the reliability of the analysis results.

[0166] Output result: The pre - processed high - quality dataset, laying a solid foundation for subsequent analysis.

[0167] 2. Key factor extraction stage (principal component and factor analysis)

[0168] Analysis object: The pre - processed dataset (for small samples).

[0169] Analysis method: Use principal component analysis and factor analysis techniques to identify and extract the key evaluation factors that affect online reputation.

[0170] Output result: A set of key evaluation factors, providing an important basis for subsequent cluster analysis.

[0171] 3. Cluster analysis stage

[0172] Analysis basis: The set of key evaluation factors.

[0173] Analysis means: Adopt a clustering algorithm to conduct a detailed grouping of online reputation and reveal the reputation characteristics of different groups.

[0174] Output result: The clustered reputation categories, providing an intuitive reference for efficiency evaluation.

[0175] 4. Efficiency evaluation stage (DEA efficiency model)

[0176] Evaluation indicators: The input indicators are all evaluation influencing factors, and the output indicators are the key indicators representing online reputation.

[0177] Evaluation method: Use the data envelopment analysis (DEA) model to evaluate the contribution degree of each indicator to the overall reputation.

[0178] Output result: A set of key influencing indicators, providing key input features for constructing a prediction model.

[0179] 5. Prediction model construction stage (BP neural network)

[0180] Model input: All evaluation indicators based on a large-scale dataset.

[0181] Model training: Use the results of principal component and factor analysis under small sample data as the target output to train the BP neural network model.

[0182] Model output: The trained BP neural network prediction model, which has the ability to predict the online reputation of new data.

[0183] In a preferred embodiment, the specific method for using principal component analysis and factor analysis techniques to identify and extract the key evaluation factors that affect online reputation is as follows:

[0184] Perform principal component analysis on the preprocessed small sample dataset and calculate the covariance matrix:

[0185]

[0186] Perform eigenvalue decomposition on C:

[0187] Cv k =λ k v k

[0188] where λ k is the eigenvalue and v k is the corresponding eigenvector

[0189] Sort the eigenvalues from largest to smallest, and select the eigenvectors corresponding to the top p largest eigenvalues to form the transformation matrix V = [v1, v2, …, v p ,];

[0190] Through:

[0191] Y = ZV

[0192] Project the original data into the new low-dimensional space to obtain the principal component matrix Y, whose dimension is m × p. These principal components are the key factors affecting the network reputation;

[0193] Suppose Z can be expressed as Z = AF + ∈, where A is the factor loading matrix, F is the common factor matrix, and ∈ is the special factor matrix. Estimate the factor loading matrix A and the common factor matrix F by methods such as maximum likelihood estimation. Select the common factors with larger variance contribution rates and combine them with the principal component analysis results to obtain the key evaluation factor set K, which provides a basis for subsequent clustering analysis.

[0194] In a preferred embodiment, a clustering algorithm is used to perform a detailed grouping of the network reputation. The specific method for revealing the reputation characteristics of different groups is as follows:

[0195] Based on the key evaluation factor set K, calculate the distance between samples using the Euclidean distance. For two samples k i and k j , their Euclidean distance:

[0196]

[0197] where and are the values of the sample and at the l-th key evaluation factor respectively;

[0198] Adopt the agglomerative hierarchical clustering algorithm. Initially, each sample point is regarded as a separate class. In each step of the merging process, calculate the distance between different classes, and use the minimum distance method, that is, the distance between two classes is the distance between the two closest sample points in the two classes;

[0199] Suppose class C a and class C b ,

[0200]

[0201] Merge the two closest clusters until all sample points are merged to reach the preset number of clusters, and obtain the clustered reputation category set C clustered , providing an intuitive reference for efficiency evaluation.

[0202] In a preferred embodiment, the specific method of using the Data Envelopment Analysis (DEA) model to evaluate the contribution of each indicator to the overall reputation is as follows:

[0203] Assume that there are q decision-making units, i.e., network entities. The input indicator is the matrix X composed of all evaluation influencing factors, with a dimension of q×p, where p is the number of key evaluation factors. The output indicator is the matrix Y composed of the key indicators that represent the network reputation. output , the dimension is q×r, and r is the number of output indicators.

[0204] Using the CCR model based on constant returns to scale, for the i-th decision-making unit, a linear programming problem is constructed:

[0205]

[0206] λ j ≥0,j=1,…,q

[0207] The objective function is maxθ i , where θ i Represents the efficiency value of the i-th decision-making unit; the DEA model aims to adjust the weight variable λ j , so that the efficiency value of decision unit i is θ i Maximization; in the context of online reputation evaluation, this means finding an optimal weight distribution method to evaluate the efficiency of the network entity in utilizing input resources to obtain output;

[0208] Input constraints: Constraints Indicates that for the rth input indicator, all decision units are based on the weight λ j The sum of the weighted inputs cannot exceed the actual input value x of the i-th decision unit ir ; where q is the total number of decision making units, λ j is the weight variable of the jth decision unit, x jr is the rth input index value of the jth decision unit;

[0209] Output constraints: Constraints Indicates that for the sth output indicator, all decision units are based on the weight λ j The sum of the weighted outputs must be greater than or equal to the actual output value y of the i-th decision unit is Multiply its efficiency value θ i ; Among them, y js is the sth output indicator value of the jth decision unit;

[0210] Non - negative weight constraint: The constraint condition λ j ≥0, j = 1, …, q ensures that the weight variable λ j is non - negative; in practical terms, the weight represents the degree of contribution of other decision - making units to the efficiency evaluation of the i - th decision - making unit. Non - negative weights are reasonable because there cannot be negative contributions;

[0211] By solving the above linear programming problem, the efficiency value θ of each decision - making unit can be obtained i . The closer the efficiency value is to 1, the higher the efficiency of the decision - making unit in using input resources to obtain outputs such as online reputation, thus providing a strong basis for screening key influencing indicators and optimizing online reputation management strategies.

[0212] In a preferred embodiment, taking the principal component and factor analysis results under small - sample data as the target output, the specific method for training the BP neural network model is as follows:

[0213] Model input: All evaluation indicators based on a large - scale dataset form the input matrix I large , with a dimension of M×n, where M is the number of samples in the large - scale dataset;

[0214] Model training: Taking the principal component and factor analysis results under small - sample data, that is, the set of key evaluation factors, as the target output;

[0215] Suppose the BP neural network includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is n, the number of nodes in the hidden layer is h, and the number of nodes in the output layer is p;

[0216] The weight matrix from the input layer to the hidden layer is W ih , and the weight matrix from the hidden layer to the output layer is W ho ;

[0217] The input of the hidden - layer nodes After being processed by the activation function σ, the hidden - layer output h out = σ(h ih );

[0218] The input of the output - layer nodes The predicted output is

[0219] Through the back - propagation algorithm, calculate the error between the predicted value and the target value k:

[0220]

[0221] Adjust the weights according to the error and continuously iterate for training.

[0222] Embodiment 2

[0223] Please refer to Figure 7 , Embodiment 2 of the present invention provides a network reputation evaluation system that integrates emotional features and semantic features, including:

[0224] A data collection and preprocessing unit, which is used to collect text data related to the evaluation object, clean the original text, segment the text into words or phrases, identify stop words, and perform noise reduction processing;

[0225] An emotional feature extraction unit, which is used to determine whether a word in the text is an emotional word according to an emotional dictionary, obtain emotional degree words and negative words, determine the emotional polarity of the word, and then judge the emotional polarity of the sentence, calculate the emotional tendency of the document; further calculate the emotional value similarity of mutually related documents according to the obtained emotional tendency value;

[0226] Semantic feature extraction, which is used to convert high-frequency keywords in the preprocessed text data into word vectors through the word2vec model to realize the vector representation of text semantics;

[0227] A single-dimensional evaluation unit, which is used to perform single-dimensional evaluation in three dimensions: emotional influence, theme influence, and behavior influence;

[0228] A multi-dimensional fusion comprehensive evaluation unit, which is used to perform multi-dimensional comprehensive evaluation by integrating emotional features and semantic features. The specific steps are as follows:

[0229] Perform dimensionality reduction processing on evaluation indicators in each dimension such as emotional influence, theme influence, and behavior influence, extract principal components and factors, and simplify the evaluation model;

[0230] Through hierarchical clustering analysis, network entities are divided into different categories, and network reputation evaluation is performed according to the characteristics of different categories; from the perspectives of input and output, the efficiency of network reputation management is evaluated based on the DEA efficiency model; the weight matrix of evaluation indicators is calculated based on the BP neural network, and the network reputation is predicted and evaluated.

[0231] Embodiment 3

[0232] Embodiment 3 of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, any step of the network reputation evaluation method that integrates emotional features and semantic features can be realized.

[0233] The computer-readable storage medium may include: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store program codes.

[0234] For the introduction of the computer-readable storage medium provided in this application, please refer to the above method embodiments, and details are not described herein again.

[0235] Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A network reputation evaluation method that fuses emotional features and semantic features, characterized in that, Including: S1. Collect text data related to the evaluation object, clean the original text, segment the text into words or phrases, confirm stop words, and perform noise reduction processing; S2. Adopt a sentiment analysis method and combine it with the cosine similarity algorithm to conduct a sentiment tendency analysis on the preprocessed text data; S3. For the preprocessed text data, convert the high-frequency keywords in the text into word vectors through the word2vec model to achieve a vector representation of the text semantics; S4. Conduct single-dimensional evaluations respectively in three dimensions: sentiment impact, theme impact, and behavior impact; S5. Conduct a multi-dimensional index fusion comprehensive evaluation by integrating sentiment features and semantic features. The specific steps are as follows: Perform dimensionality reduction processing on the evaluation indicators in each dimension such as sentiment impact, theme impact, and behavior impact, extract the principal components and factors, and simplify the evaluation model; Through hierarchical clustering analysis, divide network entities into different categories, and conduct network reputation evaluation according to the characteristics of different categories; Based on the DEA efficiency model, evaluate the efficiency of network reputation management from the perspectives of input and output; Based on the BP neural network, calculate the weight matrix of the evaluation indicators and conduct a predictive evaluation of the network reputation.

2. The network reputation evaluation method integrating emotional features and semantic features according to claim 1, characterized in that The sentiment analysis method in S2 is: Judge whether the words in the text are sentiment words according to the sentiment dictionary, obtain sentiment degree words and negative words, determine the sentiment polarity of the words, and then judge the sentiment polarity of the sentence, and calculate the sentiment tendency of the document; further calculate the sentiment value similarity of the mutually related documents according to the obtained sentiment tendency value; The specific method for calculating the sentiment tendency of the document is: S θ ＝α / β*100％ Among them, α is the number of sentences with a certain type of sentiment polarity, β is the number of sentences in document S, and the proportion of a certain type of polarity sentiment is the value S of a certain type of polarity tendency of a document θ 。 3. The network reputation evaluation method integrating emotional features and semantic features according to claim 1, characterized in that In S3, the word2vec model realizes the vectorization processing of text vocabulary through Skip-gram, specifically as follows: The basic Skip-gram model is defined as p(w i |w t ); Let w be the input vector, V w be the output vector, the dictionary size be N, and the probability of correctly predicting w j through the keyword w i is as follows: where, p(w i |w t ) represents the probability of the occurrence of the word w t given the word w i ; represents the transpose of the vector V i . In mathematics and linear algebra, the transpose of a vector is an operation that converts a row vector into a column vector or a column vector into a row vector. In the context of the Skip-gram model, usually represents the vector representation of the input vector w t , while represents the vector representation of the output vector W i . The expression represents the dot product of these two vectors, that is, the sum of the products of the corresponding elements, and is used to calculate the similarity between the input vector and the output vector.

4. The network reputation evaluation method integrating emotional features and semantic features according to claim 1, characterized in that The specific method for conducting single-dimensional evaluations respectively in three dimensions: sentiment impact, theme impact, and behavior impact in S4 is: Sentiment impact evaluation: Based on sentiment analysis and the cosine similarity algorithm, calculate the sentiment tendency value and sentiment value similarity of the text, and evaluate the sentiment impact of the network reputation; Theme impact evaluation: Use the word2vec word vector processing technology and the cosine similarity algorithm to calculate the theme similarity of the text, and evaluate the theme impact of the network reputation; Behavior impact evaluation: Evaluate the behavior impact of the network reputation by statistically analyzing user behavior data such as likes, comments, and forwards.

5. The network reputation evaluation method integrating emotional features and semantic features according to claim 1, characterized in that, The specific method for simplifying the evaluation model in S5 is: Conduct principal component analysis on the preprocessed small sample data set and calculate the covariance matrix: Perform eigenvalue decomposition on C: Cv k = λ k v k where λ k is an eigenvalue and v k is the corresponding eigenvector Sort the eigenvalues from largest to smallest, and select the eigenvectors corresponding to the top p largest eigenvalues to form the transformation matrix V = [v1, v2, …, v p ,]; Through: Y = ZV Project the original data into a new low-dimensional space to obtain the principal component matrix Y, whose dimension is m×p, and these principal components are the key factors affecting the network reputation; Suppose Z can be expressed as Z = AF + ∈, where A is the factor loading matrix, F is the common factor matrix, and ∈ is the special factor matrix; estimate the factor loading matrix A and the common factor matrix F through methods such as maximum likelihood estimation; select the common factors with larger variance contribution rates and combine them with the results of principal component analysis to obtain the key evaluation factor set K, which provides a basis for subsequent clustering analysis.

6. The network reputation evaluation method integrating emotional features and semantic features according to claim 1, characterized in that, In S5, through hierarchical clustering analysis, network entities are divided into different categories, and the specific method for network reputation evaluation according to the characteristics of different categories is as follows: Based on the set of key evaluation factors K, the Euclidean distance is used to calculate the distance between samples; for two samples k i and k j , their Euclidean distance is: Among them, and are the sample and the value on the l-th key evaluation factor respectively; The agglomerative hierarchical clustering algorithm is adopted. Initially, each sample point is taken as a separate class; in each step of the merging process, the distances between different classes are calculated, and the minimum distance method is used, that is, the distance between two classes is the distance between the two closest sample points in the two classes. Let class C a and class C b , Merge the two closest classes until all sample points are merged to reach the preset number of clusters, and obtain the set C of reputation classes after clustering clustered , providing an intuitive reference for efficiency evaluation.

7. The network reputation evaluation method integrating emotional features and semantic features according to claim 1, characterized in that The specific method for evaluating the efficiency of network reputation management from the perspectives of input and output based on the DEA efficiency model is as follows: Suppose there are a total of q decision-making units, that is, network entities. The input index is the matrix X composed of all evaluation influencing factors, with a dimension of q×p, where p is the number of key evaluation factors. The output index is the matrix Y composed of key indicators representing network reputation output , with a dimension of q×r, where r is the number of output indicators; The CCR model with constant returns to scale is adopted. For the i-th decision-making unit, a linear programming problem is constructed: The objective function is maxθ i , where θ i represents the efficiency value of the i-th decision-making unit; the DEA model aims to maximize the efficiency value θ j of the decision-making unit i by adjusting the weight variable λ i ; in the scenario of online reputation evaluation, this means finding an optimal weight allocation method to evaluate the efficiency of the network entity in using input resources to obtain outputs; Input constraint: Constraint condition It means that for the r-th input index, the weighted sum of inputs of all decision-making units according to the weight λ j shall not exceed the actual input value x of the i-th decision-making unit ir ; where q is the total number of decision-making units, and λ j is the weight variable of the j-th decision-making unit, and x jr is the r-th input index value of the j-th decision-making unit; Output constraint: Constraint condition It means that for the s-th output index, the weighted output sum of all decision-making units according to the weight λ j should be greater than or equal to the actual output value y of the i-th decision-making unit is multiplied by its efficiency value θ i ; where y js is the s-th output index value of the j-th decision-making unit; Non - negative weight constraint: The constraint condition λ j ≥ 0, j = 1, …, q ensures that the weight variable λ j is non - negative; in practical terms, the weight represents the degree of contribution of other decision - making units to the efficiency evaluation of the i - th decision - making unit. Non - negative weights are reasonable because negative contributions should not occur; By solving the above linear programming problem, the efficiency value θ of each decision-making unit can be obtained. i The closer the efficiency value is to 1, the higher the efficiency of the decision-making unit in using input resources to obtain outputs such as network reputation, thus providing a strong basis for screening key influencing indicators and optimizing network reputation management strategies.

8. The network reputation evaluation method integrating emotional features and semantic features according to claim 1, characterized in that, The specific method for calculating the weight matrix of evaluation indicators based on the BP neural network and predicting and evaluating network reputation is as follows: Model input: All evaluation metrics based on a large-scale dataset form an input matrix I large , with a dimension of M×n, where M is the number of samples in the large-scale dataset; Model training: Using the results of principal component and factor analysis under small sample data, that is, the set of key evaluation factors as the target output; Suppose the BP neural network includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is n, the number of nodes in the hidden layer is h, and the number of nodes in the output layer is p; The weight matrix from the input layer to the hidden layer is W ih , and the weight matrix from the hidden layer to the output layer is W ho ; Input of the hidden layer node After being processed by the activation function σ, the output h of the hidden layer is obtained out = σ(h ih ); Input of the output layer node The predicted output is Calculate the predicted value through the backpropagation algorithm The error between the predicted value and the target value k: Adjust the weights according to the error and continuously iterate and train.

9. A network reputation evaluation system that integrates emotional features and semantic features, characterized in that, Including: A data collection and preprocessing unit for collecting text data related to the evaluation object, cleaning the original text, splitting the text into words or phrases, identifying stop words, and performing noise reduction processing; An emotional feature extraction unit for performing emotional tendency analysis on the preprocessed text data by using emotional analysis methods and combining the cosine similarity algorithm; Semantic feature extraction for converting high-frequency keywords in the text into word vectors through the word2vec model for the preprocessed text data to realize the vector representation of text semantics; A single-dimensional evaluation unit for performing single-dimensional evaluations in three dimensions: emotional influence, theme influence, and behavior influence; A multi-dimensional fusion comprehensive evaluation unit for performing multi-dimensional comprehensive evaluations by fusing emotional features and semantic features. The specific steps are as follows: Perform dimensionality reduction processing on the evaluation indicators in each dimension such as emotional influence, theme influence, and behavior influence, extract the principal components and factors, and simplify the evaluation model; Through hierarchical clustering analysis, network entities are divided into different categories, and network reputation evaluation is carried out according to the characteristics of different categories; Evaluate the efficiency of network reputation management from the perspectives of input and output based on the DEA efficiency model; Calculate the weight matrix of evaluation indicators based on the BP neural network and perform predictive evaluation of network reputation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor to implement the network reputation evaluation method for fusing emotional features and semantic features according to any one of claims 1-8.

Images